USF Libraries
USF Digital Collections

Knowledge guided processing of magnetic resonance images of the brain

MISSING IMAGE

Material Information

Title:
Knowledge guided processing of magnetic resonance images of the brain
Physical Description:
Book
Language:
English
Creator:
Clark, Matthew C
Publisher:
University of South Florida
Place of Publication:
Tampa, Fla.
Publication Date:

Subjects

Subjects / Keywords:
Brain -- Magnetic resonance imaging   ( lcsh )
Artificial intelligence -- Medical applications   ( lcsh )
Expert systems (Computer science)   ( lcsh )
brain imaging
Dissertations, Academic -- Computer Science and Engineering -- Doctoral -- USF   ( lcsh )
Genre:
government publication (state, provincial, terriorial, dependent)   ( marcgt )
bibliography   ( marcgt )
theses   ( marcgt )
non-fiction   ( marcgt )

Notes

Summary:
ABSTRACT: This dissertation presents a knowledge-guided expert system that is capable of applying routines for multispectral analysis, (un)supervised clustering, and basic image processing to automatically detect and segment brain tissue abnormalities, and then label glioblastoma-multiforme brain tumors in magnetic resonance volumes of the human brain. The magnetic resonance images used here consist of three feature images (T1-weighted, proton density, T2-weighted) and the system is designed to be independent of a particular scanning protocol. Separate, but contiguous 2D slices in the transaxial plane form a brain volume. This allows complete tumor volumes to be measured and if repeat scans are taken over time, the system may be used to monitor tumor response to past treatments and aid in the planning of future treatment. Furthermore, once processing begins, the system is completely unsupervised, thus avoiding the problems of human variability found in supervised segmentation efforts.Each slice is initially segmented by an unsupervised fuzzy c-means algorithm. The segmented image, along with its respective cluster centers, is then analyzed by a rule-based expert system which iteratively locates tissues of interest based on the hierarchy of cluster centers in feature space. Model-based recognition techniques analyze tissues of interest by searching for expected characteristics and comparing those found with previously defined qualitative models. Normal/abnormal classification is performed through a default reasoning method: if a significant model deviation is found, the slice is considered abnormal. Otherwise, the slice is considered normal. Tumor segmentation in abnormal slices begins with multispectral histogram analysis and thresholding to separate suspected tumor from the rest of the intra-cranial region. The tumor is then refined with a variant of seed growing, followed by spatial component analysis and a final thresholding step to remove non-tumor pixels. The knowledge used in this system was extracted from general principles of magnetic resonance imaging, the distributions of individual voxels and cluster centers in feature space, and anatomical information. Knowledge is used both for single slice processing and information propagation between slices. A standard rule-based expert system shell (CLIPS) was modified to include the multispectral analysis, clustering, and image processing tools.A total of sixty-three volume data sets from eight patients and seventeen volunteers (four with and thirteen without gadolinium enhancement) were acquired from a single magnetic resonance imaging system with slightly varying scanning protocols were available for processing. All volumes were processed for normal/abnormal classification. Tumor segmentation was performed on the abnormal slices and the results were compared with a radiologist-labeled 'ground truth' tumor volume and tumor segmentations created by applying supervised k-nearest neighbors, a partially supervised variant of the fuzzy c-means clustering algorithm, and a commercially available seed growing package. The results of the developed automatic system generally correspond well to ground truth, both on a per slice basis and more importantly in tracking total tumor volume during treatment over time.
Bibliography:
Includes bibliographical references.
System Details:
System requirements: World Wide Web browser and PDF reader.
System Details:
Mode of access: World Wide Web.
Statement of Responsibility:
by Matthew C. Clark.
General Note:
Includes vita.
General Note:
Title from PDF of title page.
General Note:
Document formatted into pages; contains 222 pages.

Record Information

Source Institution:
University of South Florida Library
Holding Location:
University of South Florida
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
oclc - 40069736
notis - AJJ0728
usfldc doi - E14-SFE0000001
usfldc handle - e14.1
System ID:
SFS0024692:00001


This item is only available as the following downloads:


Full Text
xml version 1.0 encoding UTF-8 standalone no
record xmlns http:www.loc.govMARC21slim xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.loc.govstandardsmarcxmlschemaMARC21slim.xsd
leader nam 2200409Ka 4500
controlfield tag 006 m d
007 cr bn
008 030703s2001 flu sbm s000|0 eng d
datafield ind1 8 ind2 024
subfield code a E14-SFE0000001
035
(OCoLC)40069736
9
AJJ0728
b SE
040
FHM
c FHM
049
FHME
090
TK7885.Z9
1 100
Clark, Matthew C.
0 245
Knowledge guided processing of magnetic resonance images of the brain
h /
by Matthew C. Clark.
260
[Tampa, Fla.] :
University of South Florida,
2001.
500
Includes vita.
538
System requirements: World Wide Web browser and PDF reader.
Mode of access: World Wide Web.
Title from PDF of title page.
Document formatted into pages; contains 222 pages.
520
ABSTRACT: This dissertation presents a knowledge-guided expert system that is capable of applying routines for multispectral analysis, (un)supervised clustering, and basic image processing to automatically detect and segment brain tissue abnormalities, and then label glioblastoma-multiforme brain tumors in magnetic resonance volumes of the human brain. The magnetic resonance images used here consist of three feature images (T1-weighted, proton density, T2-weighted) and the system is designed to be independent of a particular scanning protocol. Separate, but contiguous 2D slices in the transaxial plane form a brain volume. This allows complete tumor volumes to be measured and if repeat scans are taken over time, the system may be used to monitor tumor response to past treatments and aid in the planning of future treatment. Furthermore, once processing begins, the system is completely unsupervised, thus avoiding the problems of human variability found in supervised segmentation efforts.Each slice is initially segmented by an unsupervised fuzzy c-means algorithm. The segmented image, along with its respective cluster centers, is then analyzed by a rule-based expert system which iteratively locates tissues of interest based on the hierarchy of cluster centers in feature space. Model-based recognition techniques analyze tissues of interest by searching for expected characteristics and comparing those found with previously defined qualitative models. Normal/abnormal classification is performed through a default reasoning method: if a significant model deviation is found, the slice is considered abnormal. Otherwise, the slice is considered normal. Tumor segmentation in abnormal slices begins with multispectral histogram analysis and thresholding to separate suspected tumor from the rest of the intra-cranial region. The tumor is then refined with a variant of seed growing, followed by spatial component analysis and a final thresholding step to remove non-tumor pixels. The knowledge used in this system was extracted from general principles of magnetic resonance imaging, the distributions of individual voxels and cluster centers in feature space, and anatomical information. Knowledge is used both for single slice processing and information propagation between slices. A standard rule-based expert system shell (CLIPS) was modified to include the multispectral analysis, clustering, and image processing tools.A total of sixty-three volume data sets from eight patients and seventeen volunteers (four with and thirteen without gadolinium enhancement) were acquired from a single magnetic resonance imaging system with slightly varying scanning protocols were available for processing. All volumes were processed for normal/abnormal classification. Tumor segmentation was performed on the abnormal slices and the results were compared with a radiologist-labeled 'ground truth' tumor volume and tumor segmentations created by applying supervised k-nearest neighbors, a partially supervised variant of the fuzzy c-means clustering algorithm, and a commercially available seed growing package. The results of the developed automatic system generally correspond well to ground truth, both on a per slice basis and more importantly in tracking total tumor volume during treatment over time.
504
Includes bibliographical references.
516
Text (Electronic thesis) in PDF format.
590
Co-adviser: Goldgof, Dmitry B.
Co-adviser: Hall, Lawrence O.
650
Brain
x Magnetic resonance imaging.
Artificial intelligence
Medical applications.
Expert systems (Computer science).
653
brain imaging.
690
Dissertations, Academic
z USF
Computer Science and Engineering
Doctoral.
949
FTS
SFERS
ETD
TK7885.Z9 (ONLINE)
773
t USF Electronic Theses and Dissertations.
4 856
u http://digital.lib.usf.edu/?e14.1



PAGE 1

GraduateSc hool Univ ersit yofSouthFlorida T ampa,Florida CER TIFICA TEOFAPPR O V AL Ph.D.Dissertation ThisistocertifythatthePh.D.Dissertationof MA TTHEWC.CLARK withamajorinComputerScienceandEngineeringhasbeenappro v edb y theExaminingCommitteeonDecem ber12,1997 assatisfactoryforthedissertationrequiremen t fortheDoctorofPhilosoph yinComputerScienceandEngineeringdegree ExaminingCommittee: Co-MajorProfessor:La wrenceO.Hall,Ph.D. Co-MajorProfessor:DmitryB.Goldgof,Ph.D. Mem ber:SridharMahadev an,Ph.D. Mem ber:LaurenceClark e,Ph.D. Mem ber:An thon yLlew elyn,Ph.D.

PAGE 2

c MatthewC.Clark 1998 AllRigh tsReserv ed

PAGE 3

KNO WLEDGE-GUIDEDPR OCESSINGOFMA GNETIC RESONANCEIMA GESOFTHEBRAIN b y MA TTHEWC.CLARK Adissertationsubmittedinpartialfulllmen t oftherequiremen tsforthedegreeof DoctorofPhilosoph yinComputerScienceandEngineering Departmen tofComputerScienceandEngineering CollegeofEngineering Univ ersit yofSouthFlorida Ma y1998 Co-MajorProfessor:La wrenceO.Hall,Ph.D. Co-MajorProfessor:DmitryB.Goldgof,Ph.D.

PAGE 4

DEDICA TION T om yGrandfatherClark,ateac her.

PAGE 5

A CKNO WLEDGMENTS Iameternallyindebtedtom yLordandSa viour,throughwhomallthingsw ere madepossibleandwhoblessedmewiththeabilit ytorealizethisw ork.Iamequally thankfultom ymother,father,andsisterwholledaplacecalledhome. Iw ouldlik etothankm yadvisorsDr.La wrenceHallandDr.DmitryGoldgof forencouragingmetopursueagoalIhadnev erconsideredandgivingmetheopportunit ytodoso.Theirsupportandwisecounselw erein v aluable.Im ustalsothank m ycommitteemem bers,Dr. LaurenceClark e,Dr. SridharMahadev an,andDr. An thon yLlew elynfortakingthetimeandeorttoensurem yw orkw asofqualit y Myresearc hw ouldha v ebeenfarmoredicultw ereitnotforDr. Robert V elth uizen,whohelpedmewithproblemslargeandsmall.Iw ouldalsolik etothank Dr.F.ReedMurtaughforexplainingthe\m ysteries"orMRimaging.Im ustalso thankLynnandMik eHeathforbeingsuc hgoodfriends,bothforhelpingmeinside thelab,andremindingmethattherew asstilllifeoutsideofit.Im ustalsothankm y friendsMik e,Nell,Russ,andMic hellefork eepingmesanewhenthec hallengeshere sometimesseemedalittleo v erwhelming. Thisresearc hw aspartiallysupportedb yagran tfromtheWhitak erfoundation andagran tfromtheNationalCancerInstitute(CA59425-01). Imagedataand additionalcomputersupportw aspro videdb ytheDigitalMedicalImagingProgram andtheMottCancerCen ter.ThankstoDr.MohanV aidy anathanforhisassistance inthegroundtruthw ork.

PAGE 6

T ABLEOFCONTENTS LISTOFT ABLES v LISTOFFIGURES vi ABSTRA CT viii CHAPTER1. INTR ODUCTION 1 CHAPTER2. BA CK GR OUND 8 2.1 SlicesofIn terestfortheStudy :::::::::::::::::::: 8 2.2 BasicMRCon trastPrinciples :::::::::::::::::::: 12 2.3 Qualitativ eModelingandAnatomicalKno wledge ::::::::: 14 2.4 Kno wledge-BasedSystems :::::::::::::::::::::: 19 2.4.1 ReasoningByDefault ::::::::::::::::::::: 20 2.4.2 RuleBasedSystems :::::::::::::::::::::: 21 2.5 F uzzySetsandthec-MeansClusteringAlgorithms ::::::::: 22 2.6 Kno wledgePropagation :::::::::::::::::::::::: 25 2.7 Relev an tImageProcessingT ec hniques ::::::::::::::: 26 2.7.1 MorphologicalOperators ::::::::::::::::::: 26 2.7.2 ImageReection ::::::::::::::::::::::: 27 2.7.3 ConnectedComponen ts :::::::::::::::::::: 28 2.7.4 MedianFilter ::::::::::::::::::::::::: 29 2.7.5 Thresholding ::::::::::::::::::::::::: 30 2.7.6 Bi-orthogonalThic kness ::::::::::::::::::: 31 2.8 SystemOv erview ::::::::::::::::::::::::::: 31 i

PAGE 7

2.8.1 P athologyDetectioninSlicesBelo wtheV en tricles ::::: 32 2.8.2 P athologyDetectioninSlicesAbo v etheV en tricles ::::: 34 2.8.3 T umorSegmen tation ::::::::::::::::::::: 36 CHAPTER3. RELA TEDW ORK 39 3.1 SingleCon trastSegmen tationMethods ::::::::::::::: 40 3.1.1 Thresholding ::::::::::::::::::::::::: 40 3.1.2 T extureAnalysis ::::::::::::::::::::::: 44 3.1.3 EdgeDetection :::::::::::::::::::::::: 46 3.2 MultispectralSegmen tationMethods :::::::::::::::: 47 3.2.1 ClusteringandNonparametricP atternRecognition :::: 48 3.2.2 Kno wledge-BasedSegmen tation ::::::::::::::: 51 3.2.3 Statistical/P arametricMethods ::::::::::::::: 54 3.2.4 NeuralNet w orks ::::::::::::::::::::::: 58 CHAPTER4. DETECTINGP A THOLOGYINSLICESBELO WTHEVENTRICLES 62 4.1 StageZero:InitialSegmen tation ::::::::::::::::::: 63 4.2 StageOne:CreatingtheIn tra-CranialMask :::::::::::: 64 4.2.1 ExtraandIn tra-CranialClusterSeparation ::::::::: 65 4.2.2 Reco v eringIn tra-CranialClustersinT emplate5LSlices :: 67 4.2.3 DetectingF alseWhiteMatter :::::::::::::::: 68 4.2.4 ExtractingtheIn tra-CranialRegion ::::::::::::: 70 4.3 StageTw o:SliceT emplateIden tication :::::::::::::: 71 4.3.1 InitialT emplateDetermination ::::::::::::::: 71 4.3.2 In tra-CranialMaskRenemen t ::::::::::::::: 76 4.3.3 T emplateV ericationwithV en tricleShapeInformation :: 82 4.3.3.1 WhiteMatterSplitting ::::::::::::::::: 82 ii

PAGE 8

4.3.3.2 Chec kingtheShapeoftheV en tricleArea :::::: 87 4.4 StageThree:P athologyDetection :::::::::::::::::: 90 4.4.1 Chec kingMaskSymmetry :::::::::::::::::: 91 4.4.2 Chec kingClusterSymmetry ::::::::::::::::: 93 4.4.3 DetectingEnhancingPixels ::::::::::::::::: 93 4.4.3.1 MultispectralHistogramThresholding :::::::: 95 4.4.3.2 RegionAnalysisofAreasofEnhancemen t :::::: 98 CHAPTER5. TUMORSEGMENT A TION 104 5.1 StageOne:Reco v eringLostT umorintheIn tra-CranialMask :: 104 5.1.1 T umorReco v eryinUpperSlices ::::::::::::::: 105 5.1.2 T umorReco v eryinLo w erSlices ::::::::::::::: 107 5.2 StageTw o:MultispectralHistogramThresholding ::::::::: 108 5.3 StageThree:\Densit yScreening"inF eatureSpace :::::::: 115 5.4 StageF our:RegionAnalysis ::::::::::::::::::::: 118 5.4.1 Remo vingExtra-CranialRegions :::::::::::::: 119 5.4.1.1 ProcessingLo w erSlices :::::::::::::::: 119 5.4.1.2 Remo vingMeningialRegionsinAllSlices :::::: 122 5.4.2 Remo vingIn tra-CranialNon-T umorRegions :::::::: 123 5.5 StageFiv e:FinalT1Thresholding ::::::::::::::::: 128 5.5.1 ThresholdingIf Enhance T 1 L=T 1 : 0 :::::::::::::: 133 5.5.2 ThresholdingIf Enhance T 1 L=T < 1 : 0 :::::::::::::: 137 5.6 P ostProcessing:T umorV ericationinUpperSlices :::::::: 140 CHAPTER6. RESUL TS 141 6.1 T rainingandT estSets :::::::::::::::::::::::: 141 6.2 ResultsforP athologyDetection ::::::::::::::::::: 145 6.3 ResultsforT umorSegmen tation ::::::::::::::::::: 149 iii

PAGE 9

6.3.1 Kno wledge-Basedvs.GroundT ruth ::::::::::::: 149 6.3.2 Kno wledge-Basedvs.SupervisedMethods ::::::::: 153 6.3.3 Ev aluationOv erRepeatScans :::::::::::::::: 159 6.3.4 Kno wledge-Basedvs.OtherEorts ::::::::::::: 166 CHAPTER7. SUMMAR Y 169 7.1 SystemSummary ::::::::::::::::::::::::::: 169 7.2 F utureW ork :::::::::::::::::::::::::::::: 172 LISTOFREFERENCES 175 APPENDICES 185 APPENDIXAMRD A T ASETPR OTOCOLS 186 APPENDIXBTUMORSEGMENT A TIONRESUL TS 189 APPENDIXCEFF OR TSUSINGEDGEDETECTION 207 APPENDIXDR ULESINCLIPS 212 VIT A 213 iv

PAGE 10

LISTOFT ABLES T able1. SummaryofMRDataAv ailable ::::::::::::::::::: 12 T able2. ASynopsisofT1,PD,andT2EectsontheMagneticResonance Image :::::::::::::::::::::::::::::::::: 13 T able3. RulesforClusterSeparation ::::::::::::::::::::: 67 T able4. ClusterCen terDistributionsofNormalandAbnormalTissues :: 95 T able5. RegionLabelingRulesBasedonPixelPresence ::::::::::: 126 T able6. Remo vingNon-T umorRegionsBasedonStatisticalMeasuremen ts 127 T able7. CandidateThresholds ::::::::::::::::::::::::: 132 T able8. SummaryofMRDataAv ailable ::::::::::::::::::: 142 T able9. NormalMRSliceDistribution :::::::::::::::::::: 143 T able10. AbnormalMRSliceDistribution ::::::::::::::::::: 144 T able11. UnseenMRSliceDistribution :::::::::::::::::::: 145 T able12. SystemF ailurestoDetectT umorinLo w erSlices. :::::::::: 147 T able13. F ailuresintheP ost-ProcessingStageforUpperSlices. ::::::: 148 T able14. Kno wledge-BasedT umorvs.RadiologistLabeledT umorP erV olume.151 T able15. kNNT umorvs.RadiologistLabeledT umorP erV olume. ::::: 155 T able16. Kno wledge-BasedT umorvs.kNN :::::::::::::::::: 157 T able17. Kno wledge-BasedT umorvs.ssF CMandISG :::::::::::: 158 T able18. FinalThresholding:\StageFiv e"vs.F uzzy-EdgeBased :::::: 211 v

PAGE 11

LISTOFFIGURES Figure1. SystemOrganizationoftheKno wledge-GuidedSystem :::::: 4 Figure2. TheThreeMajorMRImagingPlanes :::::::::::::::: 9 Figure3. SlicesofIn terest :::::::::::::::::::::::::::: 10 Figure4. V olumesoftheBrain ::::::::::::::::::::::::: 11 Figure5. T emplatesforQualitativ eTissueModels. :::::::::::::: 15 Figure6. F uzzyMem bershipforNum bersClosetoT en. ::::::::::: 22 Figure7. ConnectedComponen tsinImageSpace ::::::::::::::: 28 Figure8. Bi-orthogonalThic kness. ::::::::::::::::::::::: 31 Figure9. Ov erviewofP athologyDetectionforSlicesBelo wtheV en tricles. : 33 Figure10.Ov erviewofP athologyDetectionforSlicesAbo v etheV en tricles. : 35 Figure11.Ov erviewoftheT umorSegmen tationSystem. ::::::::::: 37 Figure12.ClassCen tersforanAbnormalSlice ::::::::::::::::: 64 Figure13.SeparatingExtraandIn tra-CranialClusters :::::::::::: 65 Figure14.Reco v eringIn tra-CranialClusters :::::::::::::::::: 68 Figure15.CreatingtheIn tra-CranialMask ::::::::::::::::::: 70 Figure16.BoundingBo xes :::::::::::::::::::::::::::: 72 Figure17.DetectingT emplate5LSlices ::::::::::::::::::::: 74 Figure18.ReningtheIn tra-CranialMask ::::::::::::::::::: 76 Figure19.Setting Threshold y :::::::::::::::::::::::::: 78 Figure20.Remo vingF alseIn tra-CranialRegions :::::::::::::::: 80 Figure21.Tw oNeigh boringClassesinTw oSlices. ::::::::::::::: 83 Figure22.ShapeAnalysistoDetectWhiteMatterSplitting. ::::::::: 86 vi

PAGE 12

Figure23.ExaminingtheV en tricleAreawithEmitter\Lines" :::::::: 88 Figure24.Chec kingIn tra-cranialMaskSymmetry ::::::::::::::: 92 Figure25.SliceswithAbnormalClusterSymmetry ::::::::::::::: 94 Figure26.DetectingP athologyUsingHistogramDistributions :::::::: 97 Figure27.IsolatingP athologyThroughThresholding ::::::::::::: 98 Figure28.EnhancingT emporalV en tricles :::::::::::::::::::: 101 Figure29.ReclaimingLostT umorPixels :::::::::::::::::::: 105 Figure30.DistributionofIn tra-CranialPixels ::::::::::::::::: 109 Figure31.HistogramDistributionsforT umorandtheIn tra-CranialRegion : 110 Figure32. J m V ersustheNum berofErrorsfortheIrisSet :::::::::: 112 Figure33.MultispectralHistogramThresholdingofFigure31 ::::::::: 114 Figure34.Densit yScreeningInitialT umorSegmen tationF romFigure33(c). 115 Figure35.UsingRegionsforT umorSegmen tation ::::::::::::::: 119 Figure36.Remo vingtheEy es :::::::::::::::::::::::::: 120 Figure37.Remo vingMeningialPixels :::::::::::::::::::::: 123 Figure38.UsingPixelCoun tstoRemo v eNon-T umorousRegions ::::::: 125 Figure39.FinalT1Thresholding :::::::::::::::::::::::: 130 Figure40.F ailuresonP athologyDetection ::::::::::::::::::: 146 Figure41.AF ailureofT umorSegmen tation :::::::::::::::::: 150 Figure42.Kno wledge-BasedT umorSegmen tationvs.GroundT ruth ::::: 154 Figure43.T rac kingT umorResponseOv erRepeatScans,P atien ts1and2 :: 161 Figure44.T rac kingT umorResponseOv erRepeatScans,P atien ts3and4 :: 162 Figure45.T rac kingT umorResponseOv erRepeatScans,P atien ts5and6 :: 163 Figure46.T umorOv erestimation :::::::::::::::::::::::: 164 Figure47.DetectingEdgesAlongT umorBoundaries :::::::::::::: 209 Figure48.EdgeStructuresforF uzzyEdgeDetection :::::::::::::: 210 vii

PAGE 13

KNO WLEDGE-GUIDEDPR OCESSINGOFMA GNETIC RESONANCEIMA GESOFTHEBRAIN b y MA TTHEWC.CLARK AnAbstract Ofadissertationsubmittedinpartialfulllmen t oftherequiremen tsforthedegreeof DoctorofPhilosoph yinComputerScienceandEngineering Departmen tofComputerScienceandEngineering CollegeofEngineering Univ ersit yofSouthFlorida Ma y1998 Co-MajorProfessor:La wrenceO.Hall,Ph.D. Co-MajorProfessor:DmitryB.Goldgof,Ph.D. viii

PAGE 14

Thisdissertationpresen tsakno wledge-guidedexpertsystemthatiscapable ofapplyingroutinesform ultispectralanalysis,(un)supervisedclustering,andbasic imageprocessingtoautomaticallydetectandsegmen tbraintissueabnormalities,and thenlabelglioblastoma-m ultiform ebraintumorsinmagneticresonancev olumesof theh umanbrain.Themagneticresonanceimagesusedhereconsistofthreefeature images(T1-w eigh ted,protondensit y ,T2-w eigh ted)andthesystemisdesignedtobe independen tofaparticularscanningprotocol. Separate,butcon tiguous2Dslices inthetransaxialplaneformabrainv olume.Thisallo wscompletetumorv olumes tobemeasuredandifrepeatscansaretak eno v ertime,thesystemma ybeusedto monitortumorresponsetopasttreatmen tsandaidintheplanningoffuturetreatmen t.F urthermore,onceprocessingbegins,thesystemiscompletelyunsupervised, th usa v oidingtheproblemsofh umanv ariabilit yfoundinsupervisedsegmen tation eorts. Eac hsliceisinitiallysegmen tedb yanunsupervisedfuzzyc-meansalgorithm. Thesegmen tedimage,alongwithitsrespectiv eclustercen ters,isthenanalyzedb y arule-basedexpertsystemwhic hiterativ elylocatestissuesofin terestbasedonthe hierarc h yofclustercen tersinfeaturespace.Model-basedrecognitiontec hniquesanalyzetissuesofin terestb ysearc hingforexpectedc haracteristicsandcomparingthose foundwithpreviouslydenedqualitativ emodels.Normal/abnormalclassicationis performedthroughadefaultreasoningmethod:ifasignican tmodeldeviationis found,thesliceisconsideredabnormal.Otherwise,thesliceisconsiderednormal. T umorsegmen tationinabnormalslicesbeginswithm ultispectralhistogramanalysis andthresholdingtoseparatesuspectedtumorfromtherestofthein tra-cranialregion.Thetumoristhenrenedwithav arian tofseedgro wing,follo w edb yspatial componen tanalysisandanalthresholdingsteptoremo v enon-tumorpixels. Thekno wledgeusedinthissystemw asextractedfromgeneralprinciplesof magneticresonanceimaging,thedistributionsofindividualv o xelsandclustercen ters ix

PAGE 15

infeaturespace,andanatomicalinformation.Kno wledgeisusedbothforsingleslice processingandinformationpropagationbet w eenslices.Astandardrule-basedexpert systemshell(CLIPS)w asmodiedtoincludethem ultispectralanalysis,clustering, andimageprocessingtools. Atotalofsixt y-threev olumedatasetsfromeigh tpatien tsandsev en teenv olun teers(fourwithandthirteenwithoutgadoliniumenhancemen t)w ereacquiredfrom asinglemagneticresonanceimagingsystemwithsligh tlyv aryingscanningprotocols w erea v ailableforprocessing.Allv olumesw ereprocessedfornormal/abnormalclassication.T umorsegmen tationw asperformedontheabnormalslicesandtheresults w erecomparedwitharadiologist-labeled\groundtruth"tumorv olumeandtumor segmen tationscreatedb yapplyingsupervisedk-nearestneigh bors,apartiallysupervisedv arian tofthefuzzyc-meansclusteringalgorithm,andacommerc iall ya v ailable seedgro wingpac k age.Theresultsofthedev elopedautomaticsystemgenerallycorrespondw elltogroundtruth,bothonaperslicebasisandmoreimportan tlyintrac king totaltumorv olumeduringtreatmen to v ertime. AbstractAppro v ed: Co-MajorProfessor:La wrenceO.Hall,Ph.D. Professor,Departmen tofComputer ScienceandEngineering DateAppro v ed: AbstractAppro v ed: Co-MajorProfessor:DmitryB.Goldgof,Ph.D. AssociateProfessor,Departmen tofComputer ScienceandEngineering DateAppro v ed: x

PAGE 16

1 CHAPTER1 INTR ODUCTION AccordingtotheBrainT umorSociet y ,appro ximately100,000peopleinthe UnitedStateswillbediagnosedwithaprimaryormetastaticbraintumorwithin thenext12mon ths[33]. Oneoftheprimarydiagnosticandtreatmen tev aluation toolsforbraintumorshasbeenmagneticresonance(MR)imaging.MRimaginghas becomeawidely-usedmethodofhighqualit ymedicalimaging,especiallyinbrain imagingwhereMR'ssofttissuecon trastandnon-in v asiv enessareclearadv an tages. MRimagescanalsobeusedtotrac kthesizeofabraintumorasitresponds(or doesn't)totreatmen t.Areliablemethodforsegmen tingtumorw ouldclearlybea usefultool[69,67,121]. Curren tly ,ho w ev er,thereisnomethodwidelyacceptedinclinicalpracticefor quan titatingtumorv olumesfromMRimages[83].TheEasternCooperativ eOncologygroup[34]usesanappro ximationtotumorcross-sectionalareainthesingleMR slicewiththelargestcon tiguous,w ell-denedtumor.Theseman ualmeasuremen ts, ho w ev er,ha v esho wnpoorreproducibilit yandtumorresponsecriteriabasedonthese man ualestimationsha v esho wnpoorcorrelationwithquan titativ e2Dand3Dmetrics[24]. Supervisedpatternrecognitionmethodsha v ealsosho wnproblemswith reproducibilit y ,duetothesignican tin traandin ter-observ erv ariancein troduced o v erm ultipletrialsoftrainingexampleselection[23].F urthermore,becausesupervision,suc hastheselectionoftrainingexamples,canbetimeconsumingandrequires domain\expertise"tobeeectiv e,supervisedmethodsareunsuitableforclinicaluse.

PAGE 17

2 Theselimitationssuggesttheneedforafullyautomaticmethodfortumorv olume measuremen t,notonlyfortrac kingtumorresponsetotherap y ,butinplanningfuture treatmen tasw ell[69,67,121,24]. Researc hin tothesegmen tationofbrainimages,bothsupervisedandautomatic,hasremainedlargelyexperimen talw ork,ho w ev er. Whilesomeeortsha v e beenproposedfordeterminingthev olumesofparenc h ymalbraintissues,m ultiple sclerosislesions,andtumors,mostreportsonMRsegmen tationha v eeitherdealtwith normaldatasets,orwithneuro-psyc hiatricdisorderswithMRdistributionc haracteristicssimilartonormals[23,56].Someapproac hesusedasinglecon trastimage, whileothersexploitedMRimaging'scapabilit ytoproducem ulti-dim ensionaldata throughm ultispectralanalysis[23].An um berofthesesattheUniv ersit yofSouth Floridaha v ealsoaddressedMRsegmen tation.Bensaid[3]andV elth uizen[124]eac h dev elopedmodicationstothefuzzyc-meansclusteringalgorithminanattemptto impro v ethequalit yoftissuesegmen tation. Namasiv a y am[87]usedfuzzyrulesto segmen tnormalbraintissues.Eortsb yLiin[70,71]sho w edthatacom bination ofkno wledge-basedtec hniquesandm ultispectralanalysiscoulddetectpathologyand labelnormaltissuesforav erysmallrangeofcon tiguousslicesin tersectingthev entriclesofthebrain.Thisrangew asexpandedb ythisauthor,inhisMaster'sw ork in[20,18],todetectabnormalitiesinm ultiplecon tiguousslices(abo v ethev en tricles),makinguppartialv olumesofabraindataset.Noneoftheseeorts,ho w ev er, includingthisauthor'sMaster'sw ork,ha v eaddressedthemorediculttaskofautomaticallyextractingenhancingtumorfromacompleteMRv olumeusingaxed parameterset,withoutoperatorsupervision. Incon trast,thisdissertationpresen tsakno wledge-basedparadigmthatproducestherstandonly(basedonexaminationofa v ailableliterature)unsupervised systemcapableofautomaticallydetectingabnormalitiesinanMRimageandsegmen tingandlabelingof complete glioblastoma-m ultiformetumorv olumes.F urther,

PAGE 18

3 thissystemhasbeentestedonalargen um berofunseenimageswithaxedparameter(rule)setandquan titativ elycomparedwith\groundtruth"images.Thisallo ws tumorresponsetotherap ytobetrac k edo v errepeatscansandaidradiologistsin planningsubsequen ttreatmen t.Moreimportan tly ,thesystem'sunsupervisednature a v oidstheproblemsofobserv erv ariabilit yfoundinsupervisedmethods,pro viding completereproducibilit yofresults.F urthermore,observ er-basedtrainingexamples arenotrequired,makingthesystemsuitableforclinicaluse.Andasnewdomaininformationbecomesa v ailableandeectiv eprocessingtoolsaredev eloped,theexible natureofakno wledge-basedsystemallo wsstraigh tforw ardexpansion,notonlyin to othertumort ypes,butadditionalbrainpathologies,suc hasm ultiple-sclerosislesions orpossiblyheadtrauma. A\slice"isdenedasam ultispectralMRimagerecordedatthein tersection ofasubjectbrainandaspecic2Dplanecreatedb yanMRcoil. The(4-5mm thic k)slicesconsideredherew eretak enfromthetransaxialplane,aplaneroughly perpendiculartothelongaxisoftheh umanbody[91],withaseriesofcon tiguous slicesforminganMRv olume.Onlytransaxialslicesthatin tersectthecerebrumare consideredb ythissystem.Theinitialsliceprocessedb ythissystem(alsocalledthe \cen terslice")liesappro ximately7to8cmfromthetopofthehead. Theinitial slicew asoriginallyc hosenb yLiin[70,71]asthestartingpoin tduetoitsuniform signalwithintheMRimagingcoilanditsw elldenedandrecognizableanatomical structures,suc hastheslice'sin tersectionwiththev en triclearea.Infact,an yslice inarestrictedregionthroughthev en triclescanbeusedastheinitialslice,asisthe casewiththesystempresen tedhere. Appro ximately8to9\upper"sliceslieabo v etheinitialslice,andanadditional 6to11\lo w er"slicesarefoundbelo wtheinitialslice.Onceprocessingofthestarting slicehasbeencompleted,thesystembeginsmo vingout w ardfromthev en triclesto considerotherslices,bothup w ardsto w ardsthetopofthebrainanddo wnto w ards

PAGE 19

4 Enhancing Pathology Detection Rule Base for Pathology Detection Rule Base for Tumor Segmentation Tumor Segmentation Pathology Found Pathology Not Found Unsupervised Fuzzy c-Means Clustering Final Tumor Segmentation Processing of Slice Halts Knowledge Engineering Rule-Based Expert System (CLIPS) Multispectral Analysis and Image Processing Tools General Domain Information -Discussions with Radiologists -Examination of MR Signal Characteristics and Brain Anatomy Figure1.SystemOrganizationoftheKno wledge-GuidedSystem thenec k. Thein ternalstructureofthebrainc hangesthroughoutanMRv olume, bothabo v eandbelo wthev en tricles. Becauseeac hsliceisprocessedseparately ho w ev er,ratherthanmodeltheen tirebrainasasingleen tit y ,sixdiscretequalitativ e braintissuemodels,called\templates,"aredened.Usinglo w-lev elkno wledge,these templatesqualitativ elymodelnotonlythein ternalbrainstructures,buttheshapeof thebrainitselfandtheextra-cranialtissuessurroundingthebrain,suc hastheey es anditsassociatedtissues. Theorganizationofthesystemissho wninFigure1.Domaininformationw as a v ailableintheformofgeneralprinciplesofMRimaging,discussionswithexperts,the distributionsofindividualv o xelsandclustercen tersinfeaturespace,andanatom y ofthebrain. Kno wledgeusefulforthedesiredtasksofpathologydetectionand tumorsegmen tationisextractedvia\kno wledgeengineering"andimplem en te das heuristics/rules.Kno wledgeisusedbothforsinglesliceprocessingandtopropagate informationbet w eenslices.Thesystemsdescribedherearecompletelyautomatic(no

PAGE 20

5 h umanin terv en tiononaperv olumebasis)afterarulesetw asbuiltfromasetof trainingimages. Inaddition,kno wledgeengineeringguidestheselectionofm ultispectralanalysis andimageprocessingtoolsthatbestexploitextractedheuristics. Thesetoolsare in tegratedin toarule-basedexpertsystemshell,CLIPS[99,40].Allnecessarylo w andhighlev elimageprocessingandm ultispectralanalysismodulesarewritteninC, in tegrateddirectlyin totheCLIPSshell,andcalledasactionsfromtherigh thand sidesoftherules. P athologydetectionisorganizedasfollo ws. Eac hMRsliceisinitiallysegmen tedb yanunsupervisedfuzzyc-meansclusteringalgorithm(F CM)[13,45].The segmen tedimageanditscorrespondingclustercen tersarethenpassedtoanexpert systemwhic husesmodel-basedrecognitiontec hniques[16]tolocatealandmark, calledafocus-of-atten tiontissue. Qualitativ emodelsofbraintissuesaredened accordingtotheslice'sappropriatetemplateandcomparedwiththeirrespectiv einstancesfromtheimagebeingprocessed.Ifasignican tdeviationfromthemodelis found,thesliceisclassiedasabnormal(andlatertumorsegmen tationisperformed onit).Otherwise,theexpertsystemlocatesthenextfocus-of-atten tiontissuebased onahierarc h yofexpectedtissues.Thisprocessisrepeatedun tileitheranabnormalit yisdetectedorallqualitativ etissuemodelsha v ebeenappro ximatelymatc hed.The systemwillproceedtothenextsliceandrepeattheclassicationstepsun tilallslices thatcomprisethev olumeareprocessed.Atotalof397slices(lyingbelo wthev en tricles)w erea v ailabletothepathologydetectionsystemdescribedinChapter4,with 65slicesusedastraining.Ofthe397slicesprocessed,391w erecorrectlyclassied. Ifaslicehasbeenclassiedasabnormal,itispassedtoasystemthatcan segmen tenhancingtumorfroman yslicein tersectingthecerebrum.T umorsegmentationbeginswiththeremo v alofallpixels(reallyv o xelssincethescannedslices

PAGE 21

6 ha v ethic kness)con tainingextra-cranialtissues.Theremainingpixelsformanin tracranialmaskinwhic hallsubsequen tprocessingisperformed.Anexpertsystemuses informationgatheredfromm ultispectralanalysistorstseparatesuspectedtumor fromtherestofthein tra-cranialmaskthroughaseriesofadaptiv ehistogramthresholdingsteps. F ollo wingthethresholdoperations,aseedgro wingbasedalgorithm called\densit yscreening"locatesandremo v esnon-tumorpixelsbasedontheirconcen trationinfeaturespace.Localstatisticsarethenappliedtospatiallyconnected componen tstodiscriminatetumorousregionfromnon-tumorousregions. Anal pixel-lev elthresholdisthenappliedtocompletethetumorsegmen tationprocess. Allpatien tcasesstudiedherearekno wntocon tainglioblastoma-m ultiform e tumor(basedonpathologyreports).Ofthetumort ypesthatarefoundinthebrain, glioblastoma-m ultiform esw ereaddressedrstbecauseoftheirrelativ ecompactness andtendencytoenhancew ellwithparamagneticsubstances,suc hasgadolinium.A totalof385slices,across33v olumesand8patien ts,kno wntocon taintumor,w ere processedb ythekno wledgebasedsystem.Thekno wledge-basedtumorsegmen tationsw erecomparedwithradiologist-labeled\groundtruth"images,with18ofthe 33casescapturingatleast90%ofgroundtruthtumor.F orthepurposesoftumor v olumetrac king,segmen tationsfromindividualslices(withinthesamev olume)are mergedtocalculatetotaltumorsizein3D.Thekno wledge-basedtumorsegmen tationsgenerallymatc hthegroundtruthimages,withthekno wledge-basedsystem correctlytrac kingtumorresponsein22of25transitions.Additionalcomparisonsare madewithsupervisedsegmen tationmethodsandcommerc iall ya v ailableseedgro wingpac k age. Thekno wledge-basedsystemissho wntooutperformthesemethods, bothincapturingmoreofthegroundtruthtumorasw ellasfollo wingtumorresponseo v errepeatscans.Thereproducibilit yofthekno wledge-basedsystemisalso demonstrated.

PAGE 22

7 Theremainderofthedissertationisdividedin tosixc hapters.Chapter2discussestheMRdomain,theslicesprocessed,someofthefundamen talsthiskno wledgebasedsystemisbuiltupon,andgiv esbriefo v erviewsofthepathologydetectionand tumorsegmen tationmodulesinthekno wledge-basedsystem.Chapter3reviewssome oftherelatedw orka v ailableonMRsegmen tation.Chapter4presen tsthepathology detectionsystemforslicesbelo wtheinitialsliceanddetailsthekno wledgeusedat eac hstep.Chapter5describesthemajorprocessingstagesinthetumorsegmen tation systemandalsodescribesthespecickno wledgeusedateac hstage. Thelastt w o c hapterspresen ttheexperimen talresults,ananalysisofthem,andfuturedirections forthisw ork. Appendicesincludedetailsconcerningtheexactscanningprotocols usedfortheMRv olumes,tumorsegmen tationresultsonaperslicebasis,anda listingoftherulesandimageprocessing/m ultispectralanalysismodulesusedinthis w ork.

PAGE 23

8 CHAPTER2 BA CK GR OUND 2.1 SlicesofIn terestfortheStudy Theslicesusedinthisresearc hareobtainedinthe(trans)axialplane,aplane roughlyperpendiculartothelongaxisoftheh umanbody[91].Figure2(a)sho wsthe orien tationoftheaxialplane.Anexampleofanormalaxialsliceaftersegmen tation issho wninFigures3(a)and(b).Eac hbrainsliceconsistsofthreefeatureimages: T1-w eigh ted,protondensit yw eigh ted,andT2-w eigh ted[121]. Thefeatureimages w ereacquiredinasinglescanningsessionandwithgen tlerestrain tsplacedonthe headtoa v oidproblemsofimageregistration.Thec haracteristicsofeac hfeatureare discussedinthefollo wingsection.Figures3(c)and(d)sho wthera wandsegmen ted imagesofanabnormalaxialslicethroughthev en tricles. Thelabeledtissuesof in terestare: cerebro-spinaluid(CSF)(darkgra y)andtheparenc h ymaltissues, whitematter(white)andgra ymatter(blac k). Intheabnormalslice,pathology (ligh tgra y)occupiesanareathatw ouldotherwisebelongtonormaltissues. Asstatedinthein troduction,therstsliceprocessedb ythissystemcomes fromasmallrangeofslicesin tersectingthev en tricles[102]appro ximately7to8 cmfromthetopofthehead.Thisinitialslice(alsocalledthe\cen terslice"),w as originallyc hosenb yLi[70,71]duetoitsuniformsignalwithintheMRimaging coilandw elldenedandrecognizableanatomicalstructures,assho wninFigure3.

PAGE 24

9 Coronal Axial SagitalFigure2.TheThreeMajorMRImagingPlanes.Imagesareacquiredfromtheaxial plane. Usingtheinitialsliceasapoin tofreference,theremainderofthebrainv olumeis dividedin tot w oprimarysubsets. Figure4(a)sho wstheset w osubsets: \upper" slices,whic hcon tainallslicesabo v ethev en tricles,includingtheinitialslice,and \lo w er"slices,whic hcon tainallslicesbelo wthev en tricles(initialslice). Abrain v olumema ycon tainappro ximately8to9upperslices,and6to11lo w erslices, dependingonslicethic knessandlocationoftheinitialslice.T ypicalslicethic knesses forv olumesuseinthisdissertationrangefrom4to5mm.Figure4(b)sho wsthemajor brainstructuresthatwillbereferencedwhendescribingqualitativ ebraintissuemodel (called\templates")inSection2.3. Thetumorsegmen tationsysteminChapter5 processesslicesfromeitherrange. T able1liststheMRimagev olumesusedforthisresearc h.Threet ypesofMR imagev olumesw erea v ailableforprocessing.V olun teersw erehealth ysubjectswho v olun teeredforMRscanningandreceiv ednogadoliniumenhancemen t.Gadolinium enhancednormalsw eresubjectswhow ereadministeredgadoliniumpriortoscanning,buttheMRimagerev ealednopathologicaltissues.P atien tsv olumescon tained radiologistdiagnosedglioblastoma-m ultiform etumor(GradeIVGlioma)andhad

PAGE 25

10 (a) Ventricles(b) (c) Pathology(d) Figure3.SlicesofIn terest:(a)ra wdatafromanormalslice(T1-w eigh ted,PDand T2-w eigh tedimagesfromlefttorigh t)(b)aftersegmen tation(c)ra wdata fromanabnormalslice(T1-w eigh ted,PDandT2-w eigh tedimagesfrom lefttorigh t)(d)aftersegmen tation.White=whitematter;Blac k=gra y matter;Darkgra y=cerebro-spinaluid;Ligh tgra y=pathologyin(b) and(d). receiv edv aryinglev elsoftreatmen tpriortoinitialv olumeacquisitionandbet w een subsequen tacquisitions,includingsurgery ,radiationtherap y ,andc hemo-therap y .All patien tv olumesreceiv edgadoliniumenhancemen t. F romthea v ailableslices,atrainingsetw ascreatedtoextractheuristicrules. Therulesarebasedonkno wledgedescribedinSections2.2and2.3,forpathology detectioninthe\lo w er"slicesandtumorsegmen tationthroughouttheMRv olume. T ables9and10inChapter6listthedistributionandusageofthetrainingslices. Kno wledgeextractionisnotautomated,buth umanassistedandmadeas\general" aspossibletoa v oiddependenceonaparticularslicethic kness,scanningprotocol,or lev elofsignalin tensit y .Systemgeneralit yisdiscussedinChapter7andthescanning protocolforeac hMRv olumeislistedinAppendixA.

PAGE 26

11 Cerebellum Limbic Lobe Ventricles Parietoccipital Sulcus Central Sulcus Preoccipital Notch Upper Slices Lower Slices Center Slice 3 2 1 4 5 5L Template(a)MedialViewofBrainHemisphere Parietal Lobe Temporal Lobe Occipital Lobe Frontal LobeParietoccipital Sulcus Central Sulcus Preoccipital Notch Cerebellum(b)LateralViewofBrainHemisphere Figure4.V olumesoftheBrain. Slicesabo v etheinitialslice,sho wnin(a),are referredtoas\upper"slices,whilethosebelo wtheinitialslicearecalled \lo w er"slices.Themedial(a)andthelateral(b)viewssho wthegross anatom yofthebrain. Landmarksareincludedforalignmen tpurposes. Thetemplatesmark edin(a)aredescribedinSection2.3.Thisgureis basedonanimagefrom[44].

PAGE 27

12 T able1.SummaryofMRDataAv ailable. T ype #Subjects #V olumes #Slices V olun teer 13 26 194 GadoliniumNormals 4 4 27 P atien t 8 33 417 T otal 25 63 638 2.2 BasicMRCon trastPrinciples Oneofthek eyadv an tagesinMRimagingisitsabilit ytoacquirem ultispectral datab yrescanningapatien twithdieren tcom binationsofpulsesequenceparameters (inthiscase,repetitiontime(TR)andec hotime(TE)).F orexample,asstatedin Section2.1,theMRdatausedinthisstudyconsistsofT1,protondensit y(PD),and T2-w eigh tedfeatureimages.AT1-w eigh tedimageisproducedb yarelativ elyshort TR/shortTEsequence,aPD-w eigh tedimageusesalongTR/shortTEsequence, whilealongTR/longTEsequenceproducesaT2-w eigh tedimage[78,32].F orthe purposeofbrevit y ,theT1-w eigh ted,PD-w eigh ted,andT2-w eigh tedfeatureswillbe referredtoasT1,PD,andT2respectiv ely Aparticularpulsesequenceparametersetwillpro videthebestcon trastbet w eendieren ttissuet ypes[78]inanindividualimageandaseriesoftheseimages canbecom binedtopro videam ultispectraldataset. Theexactph ysicsofthese pulsesequencesareoutsidethescopeofthisdissertationandtheirdiscussionisleft tootherliteraturesources[78,32,106]. Theprimaryconcernhereiswhic hpulse sequencebestdelineatesspecictissues.Abriefsynopsisissho wninT able2(based on[78,32,106,49]). Giv enanMRdataset,ascatterplotofindividualpixels/v o xelscanbeformed thatisconsisten twithT able2. F orexample,T able2indicatesthatparamagnetic

PAGE 28

13 T able2.ASynopsisofT1,PD,andT2EectsontheMagneticResonanceImage. TR=RepetitionTime;TE=Ec hoTime. PulseSequence Eect Tissues (TR/TE) (SignalIn tensit y) T1-w eigh ted ShortT1relaxation F at,Lipid-Con tainingMolecules, (short/short) (brigh t) ProteinaceousFluid,P aramagnetic Substances(Gadolinium) LongT1relaxation Neoplasms,Edema,CSF, (dark) PureFluid,Inammation PD-w eigh ted Highprotondensit y F at,Blood (long/short) (brigh t) Fluids,CSF Lo wprotondensit y Calcium,Air, (dark) FibrousTissue,CorticalBone T2-w eigh ted ShortT2relaxation Ironcon tainingsubstances (long/long) (dark) (blood-breakdo wnproducts) LongT2relaxation Neoplasms,Edema,CSF, (brigh t) PureFluid,Inammation substances(usedtoenhancebrainpathology)willha v eashort/brigh tT1-w eigh ted signal,whileCSFwillha v earelativ elylong/darkT1-w eigh tedsignal.Thereforepixelsbelongingtothet w orespectiv eclasseswillha v easimilardistribution. When segmen tedb yaclusteringalgorithm,theresultingclusters(andtheirrespectiv eclustercen ters)willha v easimilardistribution. This\distribution"infeaturespacehasbeenin v estigatedb yan um berofresearc hers[111,120,119,38]andformsanimportan tfoundationinthekno wledgebasedsystemhere,asw ellasw ork[70,71].Thissynopsisisalsothestartingpoin tfor acquiredkno wledge,whic hw asrenedforthespecictasksofpathologydetection andtumorsegmen tation.InChapters4and5,whenspecicprocessingstagesare presen ted,therelev an tkno wledgewillbedetailed.

PAGE 29

14 2.3 Qualitativ eModelingandAnatomicalKno wledge Whiletheinformationpresen tedinSection2.2isextremelyuseful,itdoesha v e somelimitations.First,MRdistributionscanv arybet w eenv olumes,dependingon thesubject'sageandgender,asw ellasscanningprotocolssuc hasslicethic kness andtheactualTR/TEparametersusedinimageacquisition[53].Secondly ,dieren t tissuescansometime sha v esimilarMRc haracteristics,especiallywhengadoliniumis in troduced.Thereforeha vinginformationthatisindependen toffeaturespacew ould mak ethekno wledgebasemorerobust. Anatomicalkno wledgeisusefulintestingandv erifyingclusters(fromanF CM segmen tation)thatarecandidatesforspecictissuelabels. Italsopro videsinformationforthemoregeneralproblemofnormal/abnormalclassicationthrougha \defaultreasoning"method[98]thatsearc hesforsignican tdeformationsfromexpectedqualitativ etissuemodels.Licreatedtheoriginalqualitativ emodelin[70,71] fortheinitial/cen terslice. Sincethein ternalshapeofthebrainc hangesthroughoutthev olume,ho w ev er,ratherthanattemptingthecomplextaskofmodelingthe en tirebrainasasingleen tit y ,itw asdecidedthatitw ouldbebrok enin todiscrete spatialregionsbaseduponanatomicalstructure.Aparticularinputslicecouldthen bematc hedagainsttheappropriatequalitativ emodeltodetecttissuedeformations. Figure5sho wsanormalslicefromeac hoftheprimarymodels,calledtemplates, createdfromcarefulexaminationofa v ailabletrainingv olumes.Thetopofeac himage represen tsthe\fron t"ofthebrain. Figures5(a)through(c)sho wthetemplates createdtoprocessthe\upper"slices. Themodelssho wninFigures5(d)through (f)processthe\lo w er"slices. Inallcases,theprimarytissuesarewhitematter (ligh tgra y),gra ymatter(darkgra y)andCSF(blac k).Thetemplatescreatedw ere baseduponthein ternalshapeofwhitematterandCSF,asw ellastheshapeofthe brainitself.Eac htemplatehasdistinctc haracteristics,aswillslicesthat\t"the

PAGE 30

15 (a)T emplate1 (b)T emplate2 (c)T emplate3 (d)T emplate4 (e)T emplate5 (f)T emplate5L Figure5.T emplatesforQualitativ eTissueModels.Thetopofeac himageisconsideredthe\fron t"ofthebrain,whilethebottomisthe\bac k"ofthe brain. template.\Fitting"simplymeansthataninputslice'sin tra-cranial,CSF,andwhite matterqualitativ ec haracteristicsarev erysimilar(thoughnotnecessarilyanexact matc h)tothoseofaspecicbraintemplate. Theimagessho wnforT emplates1 through4(Figures5(a)through(d))w erecompletelygeneratedb ythekno wledgebasedsystem.TheimagesforT emplates5and5L(Figures5(e)and(f))required someman ualtissuelabeling,whic hisdetailedaseac hT emplateisdescribed. AT emplate1sliceissho wninFigure5(a)andincludestheinitialslice.Itis generallydenedb y:(1)asingle,symmetric alregionofwhitematter,(2)asingle distinct\buttery"shapedv en triclearea(V A)withcon tiguouswhitematteralong bothv erticalsides,(3)CSFoccupiestheV Aandsurroundsthein tra-cranialregion alongtheperimeter,enclosinggra ymatter,(4)CSFllingtheV Aissymmetri cal alongthev erticalaxis,asareallin tra-cranialtissues(5)gra ymattersurroundswhite matterandtendstooccup ytheedgeofthein tra-cranialregion. AnMRv olume

PAGE 31

16 generallyhasonetot w oslicesabo v eandbelo wtheinitialslice,foratotalofthree tov eslices. Mo vingup w ards,aT emplate2sliceisfoundimme diatelyabo v etheT emplate 1sliceclosesttothetopofthehead. AT emplate2slice(Figure5(b))issimilar toaT emplate1slice,butincludesthebodyofthecorpuscallosium.TheV Ahas beensplitin tot w osmallero v alareasandwhitematterisbeingpartiallyseparated b ythein ter-hemispheri cssure.Inthev olumesprocessedb ythissystem,onlyone T emplate2slicehasbeenfoundperv olume. TheremainingupperslicesareconsideredT emplate3(Figure5(c)),wherethe V Ahasdisappeareden tirely ,lea vingCSFtobefoundonlyonthebrainperimeterand bet w eenthet w ohemispheres.Whitematterseparationiscompleteandsymmetry ofallin tra-cranialtissuesalongthev erticalaxisisstillin tact.Appro ximatelyv e T emplate3slicescanbefoundinanMRv olume. Inthelo w erslices,T emplate4(Figure5(d))sliceslieimme diatelybelo wthe T emplate1sliceclosesttothenec kandaredistinguishedb ytheshrink ageofthe fron tallobeandthebreakdo wnofthebutteryshapeofthewhitemattersurrounding thev en triclearea.TheCSFcon tainedinthev en triclesbeginstodisappear,though poc k etsofCSFcanbeseenformingbet w eenthefoldsofthetemporalandfron tallobes (sho wninFigure4(b)).V erticalsymmetryofthein tra-cranialtissueisstillin tact. InsomeT emplate4slices,thecerebellumbeginstoappearwithintheoccipitallobe (atthe\bac k"ofthebrain).WhileoccasionallyfoundinaT emplate1slice,theey es, ocularnerv es,andm usclesarem uc hmoreeviden tinT emplate4slices.Threetov e T emplate4slicescangenerallybefoundinanMRv olume. Thespatialarrangemen tofwhiteandgra ymatter(alsocalledtheparenc h ymal tissues)andCSFbecomesmorein tricateandm uc hlessreliableinaT emplate5slice (Figure5(e)),especiallywithinthecerebellum.T emplate5slicescanbedetected, ho w ev er,b ynotingthedisappearanceofthefron tallobeandtheextensionofthe

PAGE 32

17 temporallobes,resultingina\horseshoe"shape.Theey esandtheirassociatedtissuesarealsofoundnearthefron t-mostextremeofeac htemporallobe. Mostofa T emplate5slicebelongstothecerebellumandt w otothreeT emplate5slicesare foundinanMRv olume.TheT emplate5sliceinFigure5(e)hadasingleclustermanuallylabeledbecausewhitematter\splitting"(o v er-segmen tationduringclustering) w asnotc hec k ed.Whitemattersplittingisc hec k edonlyduringtheexaminationof theshapeofthewhitemattersurroundingtheV Afordetectingtransitionsbet w een T emplate1and4.Arulecouldbedev elopedforwhitemattersplittinginT emplate5 slices,butduetothein tricatearrangemen tbet w eenwhiteandgra ymatter,especially withinthecerebellum,additionalkno wledge(includinggroundtruth)isnecessaryfor accuratedelineation. T emplate5Lslices(theletterLstandingfor\lo w",sho wninFigure5(f)) arenotconsideredforpathologydetection,becauseunlik etheothertemplates,the assumptionofasinglein tra-cranialregionisviolatedastheremainingpartsofthe temporallobes(andcerebrum)becomevisuallyseparablefromthecerebellum.In T emplates1through5,aconsisten tpieceofanatomicalkno wledgeisthatextracranialtissues(air/bone,bonemarro w,skin,fat,m uscle)surroundthein tra-cranial region.Thiskno wledgeisuseful(Chapter4)inextractingthebrainfromtherest ofthehead.InSection5.4.1.2,itisusedtodiscriminateextra-cranialtissuesfrom enhancingtumor.Also,gliomas(ofwhic hglioblastoma-m ultiform esareasubt ype) areprimarilyfoundwithinthecerebrum,notthecerebellum[100].Th us,aT emplate 5Lsliceisprocessedfortumorsegmen tationonlyifpathologyhasbeendetectedin thesliceimme diatelyabo v eit.Lik etheT emplate5image,theT emplate5Limage inFigure5(f)hadwhitematterman uallylabeled,butthetemporallobethathad separatedfromthecerebellumw asalsoman uallyreco v eredsincethecurren tsystem assumesasinglein tra-cranialregion.

PAGE 33

18 Asmen tionedearlier,thesemodelsarediscrete.Acquiredimagesaredependen t ontheexactMRpulsesequence,thecoilloading,andthepatien t'sage,gender,and positionrelativ etothecoil,allofwhic harev olun teerorpatien tdependen t.Hence, clusteredslicesin v ariablymatc honeofthethreemodelsonaimpreciseorfuzzybasis. Thist ypeofv ariationisallo w ed,ho w ev er,duringpathologydetection. Another importan tfeatureofthesemodelsisthattheyarestrictlyorderedtoreectthe in ternalstructureofthebrain.Thisordering,mo vingdo wn w ardsfromthev en tricles, isT emplate1,4,5,and5Landisusefulkno wledge.F orexample,onceaparticular sliceispositiv elyiden tiedasT emplate4,mo vingdo wn w ard,nosucceedingslicema y beclassiedasaT emplate1.ThisoccursbecauseT emplate1liesonlyinanareaof thev olumealreadyprocessed. Thetemplatesdescribedabo v earemostusefulinestablishingqualitativ etraits usedb yareasoningb ydefaultmethod(Section2.4.1)todetectdeformationswithin anMRslice.Thenatureofglioblastoma-m ultiform etumors,ho w ev er,sev erelycomplicatestheuseofanatomicalkno wledge,sincetheycanha v ean yshapeandoccup y an yareawithinthecerebrum.Thisprev en tsqualitativ etumormodelsfrombeing usedreliably Althoughanatomicalkno wledgecannotbeeasilyappliedtomodel tumors,itcanbeusedtodiscriminateareasthatarekno wntocon tainnotumor. T able2sho wsthatcertainextra-cranialtissuessuc hasfat,willha v eabrigh t T1-w eigh tedsignalin tensit y Extra-cranialtissuesthatreceiv easignican tblood supplywillalsoreceiv egadoliniumenhancemen t,articiallybrigh teningtheirT1w eigh tedsignalin tensit y Thiscanin terferewiththeassumptioninthekno wledgebasethatareaswiththehighestT1-w eigh tedmeanv aluecon taingadoliniumenhancedtumor.Sincetheanatomicalla y outoftheseextra-cranialtissuesismore easilymodeledthantumor,ho w ev er,processingstepsfocusontheirremo v alrather thaniden tifyingtumor,assho wninChapter5.

PAGE 34

19 Lastly ,animportan tanatomicalmec hanismisfoundandexploitedregardless ofaslice'sparticulartemplate.Calledtheblood-brainbarrier(BBB),itisoneofthe mostimportan tpiecesofkno wledgefordetectingenhancing-pathology .Inanormal brain,theBBBactsasanextremelyselectiv elteringdevice,allo wingonlyalimited n um berofnaturallyoccurringsubstances,suc haso xygenandglucose,tomigratefrom thebloodsupplyin tothebrainitselfandexcludesman yothercompounds,including paramagneticsubstanceslik egadolinium.Thepresenceoftumorsandotherbrain pathologies,ho w ev er,damagebraintissuesandaltertheBBB.This\breakdo wn"of theBBBallo wsparamagneticsubstancestoen terthetumorandenhanceitinMR images[118,92].ThepathologydetectionsystemdescribedinSection4.4.3,asw ell asthetumorsegmen tationsysteminChapter5,relyhea vilyonthisfact. 2.4 Kno wledge-BasedSystems Thesystempresen tedhereisa\kno wledge-basedsystem."Kno wledgeisan y heuristicorc h unkofinformationthathelpsdiscriminateoneclasst ypefromanother[40].InthedomainofMRIv olumes,therearet w oprimarysourcesofkno wledge a v ailable.Therstispixelin tensit yinfeaturespacebasedontissuec haracteristics withintheMRimagingsystem.T able2listsexamplesofsuc hkno wledge.Thesecond isimage/anatomicalspaceandincludesexpectedshapesandplacemen tsofcertain tissueswithintheMRimage,suc hasthetemplatesdescribedinSection2.3.Inorder tobeuseful,thekno wledgem ustbeimpleme n tedinaneectiv emanner.Thissection describessomeofthetoolsusedforthispurpose.

PAGE 35

20 2.4.1 ReasoningByDefault Reasoningb ydefault[98]isadefeasiblenon-monotonicmethodofinferring conclusionswhencompletekno wledgeisuna v ailable,orc hangingthoseconclusions whennew,andcon tradictory ,informationbecomesa v ailable.Giv enah ypothesis S reasoningb ydefaultassumes S tobetrueunlessandun tilevidenceisdisco v eredthat pro v es S in v alid,inwhic hcasetheh ypothesisiswithdra wn.Moreformally: Default Sastrueunlessandun tilSisdispro v ed. Thisconceptcanbeexpandedtosimplifydecisionsandsa v espace.F orexample,whendrivingtoaparticularlocation,mostpeopletak ethesameroute,making av ariet yofassumptionsabouttra v elingconditions.Whenoneoftheseconditions isviolatedb y\exceptions,"suc hasconstruction,therouteisadjusted.Sinceitis m uc heasiertostoreasmallersetof\exceptions,"(tra v elingconditionsthatrequire thetra v elroutebec hanged)thanthem uc hlargersetofconditionswherethenormal routecanbefollo w ed,aw ork ablerulemigh tbe: Ifcurren ttra v elingconditionsdonot matc han yofthe\exceptions"storedandnoothercon traryevidenceexists,conclude thatthenormalroutema ybetak en. DetectingpathologywithinanMRslicecanbeacomplexproblem.Asstated inSection2.3,thev ariancebet w eenpatien ts,MRcoils,headpositions,scanning protocols,etc,prev en tsaquan titativ edescriptionofMRv olumes.Bycreatingqualitativ emodelsandusingreasoningb ydefault,ho w ev er,theproblemcanbesimplied b ysearc hingfordeformationsfromthesemodels.Giv enaninputslice,aninstance ofatissueismatc hedagainstitscorrespondingqualitativ emodel.Ifnosignican t deviationfromthemodelisfound,theinstanceisconsiderednormalun tilevidence tothecon traryisfound.Reasoningb ydefaultisalsousedfortumorsegmen tation. Sincenormalbraintissuescanbemorereliablymodeled,pixelsbelongingtothem aredetectedandremo v ed,lea vingtheremainingpixelstobelabeledastumor.

PAGE 36

21 2.4.2 RuleBasedSystems Arulebasedsystemconsistsofthreefundamen talelemen ts:aw orkingmemory offactsfromwhic hinferencesarederiv ed,akno wledgebaseofrulestoguidethe inferences,andtheinferenceenginetodra wconclusionsandrerules[40,108].Rules resem bleIF-THENstatemen tsandha v ethebasicform: h ANTECEDENT i = )h CONSEQUENT ( S ) i Thelefthandsideoftherule, h ANTECEDENT i ,con tainsthesetofconditions requiredfortheruletore,whiletherigh thandside, h CONSEQUENT ( S ) i ,isaset ofoneormoreresultan tactions.Whenallofthelefthandconditionsofaruleare satisedb ythefactsinw orkingmemory ,theruleisredb ytheinferenceengineand allofitsrigh thandactionsareexecuted,whic hma ymodifythew orkingmemory Whendomainspecickno wledgeisa v ailable,an\expertsystem"isoneof themostcommonimplem en tationsofrulebasedsystems. Domainkno wledgeis explicitlyin tegratedin totherulebasedsystem(storedasrules)throughaprocess called\kno wledgeengineering"[79].Kno wledgediscussedinSection2.3andlistedin T able2areexamplesthatw ereman uallyextractedandin tegratedin tothesystem. Inthisdissertation,theexpertsystemisimplem en tedthroughtheCLanguage IncorporatedProductionSystem(CLIPS).CLIPSisastandardrulebasedsystem andhasaw elldenedin terfaceprotocolwiththeClanguage. Thisallo wsadditionalfunctionstobewrittenandin tegrateddirectlyin totheCLIPSshell,including imageprocessing,patternrecognition,andm ultispectralanalysisfunctions. More informationcanbefoundin[99,40].

PAGE 37

22 A u(A)Figure6.F uzzyMem bershipforNum bersClosetoT en. 2.5 F uzzySetsandthec-MeansClusteringAlgorithms F uzzysetsallo welemen tstoha v epartialmem bershipinm ultipleclasses.In classical(crisp)logic,anelemen teitherbelongsordoesnotbelongtoaset A .Ina fuzzyset,ho w ev er,amem bershipgradeisassociatedwitheac helemen ttoreectthe degreetowhic htheelemen tbelongstoset A .Thisfuzzymem bershipcanha v eav alue an ywherebet w een0(completeexclusion)and1(completemem bership),inclusiv e.An elemen twithmem bership0.6belongsmoretoset A thananelemen twithmem bership 0.59,althoughbothbelongonlypartiallytotheset.Amem bershipgradema ybe mappedin toaqualitativ econceptinanapplication. Figure6sho wsafuzzysetdenedtodescriben um bersthatare\closetoten". Giv enn um beraA,someelemen tsoftheset(intheform( h A i ,grade ( A )))are f ( h 8 i ; 0 : 2),( h 9 i ; 0 : 5),( h 10 i ; 1 : 0),( h 11 i ; 0 : 5),( h 12 i ; 0 : 2) g .Onema yconcludethatif themem bershipgradeof h A i isnotlessthan0.5,then um beris\closetoten." Asstatedinthein troduction,therststepinprocessinganinputMRsliceis toapplythefuzzyc-meansclustering(F CM)algorithm[13,45]toac hiev eaninitial

PAGE 38

23 segmen tationofexpectedtissueswithintheMRslice.Abriefreviewofthealgorithm isgiv enasfollo ws. Giv enadatasetof p tuplesofreals f x 1 x 2 ,..., x n g ,whic histobepartitioned in to c classes,afuzzyc-meansclusteringalgorithm(F CM)creates c fuzzysetsanda setof c initialclasscen ters f v 1 v 2 ,..., v c g ,eac hbeing p tuplesofreals.The i th fuzzy sethastheform U i = f i 1 i 2 ,..., in g ,inwhic h ik represen tsthat x k belongsto the i th classwithamem bershipgrade ik .The c fuzzysetsformthero wsofafuzzy partitionmatrix U .F CMiterativ elyminimi zesthefollo wingobjectiv efunction: J m = n X k =1 c X i =1 ( ik ) m ( d ik ) m where d 2 ik = k x k )Tj/T18 1 Tf49 0 TD(v i k 2 isan yinnerproductnormmetric.Thisleadstotheclusteringofthegiv endataset in to c fuzzysets.Theiterativ eoptimizationiscarriedoutasfollo ws. 1.Initializetheclustercen termatrix V 2.Set b =0, 3.If b 6 =0calculatethe c clustercen ters v ( b ) i with U ( b ) : v ( b ) il = P n k =1 ( ( b ) ik ) m x kl P n k =1 ( ( b ) ik ) m ; l =1 ; 2 ;:::;p: 4.Update U ( b ) :fork=1ton, I k = f i j 1 i c;d ik = k x k )Tj/T18 1 Tf50 0 TD(v i k =0 g I k = f 1 ; 2 ;:::;c g)Tj/T17 1 Tf84 0 TD0 Tc(I k ; forthe k th columnofthematrix U ( b ) if I k = ; ,then ik = 1 P c j =1 ( d ik d jk ) 2 m )Tj/T9 1 Tf24.0001 0 TD(1 else,set ik =0foralli 2 I k and ik =1;

PAGE 39

24 5.Compare U ( b ) and U ( b +1) U = c X i =1 n X k =1 k ( b +1) ik )Tj/T17 1 Tf50 0 TD( ( b ) ik k if U < ,stop;otherwise,set b = b +1,gotostep3. T ostartF CMb yinitializingthemem bershipmatrix U ,steps1-3abo v ecan bemodiedasfollo ws: 1.Initializethefuzzypartitionmatrix U 2.Set b =0, 3.Calculatethe c clustercen ters v ( b ) i with U ( b ) : v ( b ) il = P n k =1 ( ( b ) ik ) m x kl P n k =1 ( ( b ) ik ) m ; l =1 ; 2 ;:::;p: Ineitherinitializationsc heme,therearesev eralparameterstobedecidedwhen usingF CM.Inthissystem,theyare:thew eigh tgiv entothedistanceandmem bership v alues m =2,thestoppingcriteria =0 : 225, k x )Tj/T18 1 Tf38.9999 0 TD(v k 2 =( x )Tj/T18 1 Tf38.9999 0 TD(v ) T ( x )Tj/T18 1 Tf39 0 TD(v ),theEuclidean distance,andthematrix V (0) whic hisinitializedasfollo ws. fori=1tos,wheresisthen um beroffeatures nd MIN i ,theminim umv alueforall x i; 1 ::n nd MAX i ,themaxim umv alueforall x i; 1 ::n end fori=1tos MIN i = MIN i 1 : 1 MAX i = MAX i = 1 : 1 dene STEP i = MAX i )Tj/T16 1 Tf26.9999 0 TD4 Tc[(MIN i c end fori=1tos v 1 ;i = STEP i

PAGE 40

25 end forj=2toc v j; 1 = MAX 1 )Tj/T20 1 Tf49 0 TD(( j )Tj/T20 1 Tf49 0 TD(1) STEP 1 fori=2tos v j;i = MIN i +( j )Tj/T20 1 Tf50 0 TD(1) STEP i end end Theminim umv alue, MIN i ,w asm ultiplie db y1 : 1andthemaxim umv alue, MAX i ,w asdividedb y1 : 1toinsurethatallinitialclustercen tersw erewithinthe inputdatarange.Aclustercen terlocatedoutsidethedatarangema yprev en tF CM fromcon v ergingtousefullocalextrema.Therstclustercen terw assetto STEP to co v erakno wnaircluster.Theremainingclustersw ereinitializedsuc hthattherst feature(T1)decreasedastheremainingfeatures(PDandT2)increased.Thisisto reectkno wledgeinSection2.2thatin tra-cranialclustersofin teresttendtodecrease intheirT1-w eigh tedcen troidv alueastheirPDandT2-w eigh tedv aluesincrease. 2.6 Kno wledgePropagation Kno wledgepropagationexpandssinglesliceprocessingin tov olume-basedeffortsb ysa vingkno wledgegainedfromprocessingasliceandusingitasaguidelinefor decisionsinprocessingadjacen tslices.F orexample,whenacurren tslice'stemplate isbeingdetermined,thetemplatemodeland,insomecases,shapeofthein tra-cranial maskoftheprevioussliceguidesthedecisionprocess,allo wingineligiblemodelsto beexcludedthroughthetemplateorderingheuristicdescribedinSection2.3. In T emplate5Lslices,thein tra-cranialmaskfromtheT emplate5slicespatiallynearest

PAGE 41

26 tothenec kisusedtoreco v erin tra-cranialclustersandguidetumorsegmen tationin abnormalcases.Also,aT emplate5Lsliceisconsideredonlyifitsimmedi ateneigh bor w asfoundtocon taintumor. 2.7 Relev an tImageProcessingT ec hniques Adatasetof p tuplesofrealsofsize n f x 1 x 2 ,..., x n g ,canbeview edas p featureimages.Eac hfeatureimagecanbetreatedasat w o-dimensionalarra y f x 1 ; 1 x 1 ; 2 ,..., x 1 ;k x 2 ; 1 x 2 ; 2 ,..., x j;k g andtheproductofthewidth j andheigh t k ofthearra yequals n .Eac himageelemen tiscalledapixel.Inthissection,there aret w oprimaryclassesofpixels:\foreground"pixelsarepixelsofin terest,while \bac kground"pixelsarenot. Inthisw ork,bac kgroundpixelsha v eav alueof0, whileforegroundpixelsreceiv eapositiv enon-zerov alue.Thesebinaryimagesare oftentreatedasimage\masks"whic hrestrictprocessingtoonlyforegroundpixels con tainedb yaparticularmask. 2.7.1 MorphologicalOperators Morphologicalprocessingreferstooperationswhereaninputobject(usually binary)ismodiedb yanotherobjectcalleda structuringelemen t torev ealamore usefulorin terestingshape[54,103,55].Thet w ofundamen taloperationsare erosion ( )and dilation ( ).Let B x denotethatthestructuringelemen tistranslatedto poin t x .Theerosionoperatorofobject X b ystructuringelemen t B x isdenedas thesetofallpoin ts x suc hthat B x isincludedin X X B = f x : B x 2 X g

PAGE 42

27 Similarly ,thedilationoperatorof X b y B x isdenedasthesetofallpoin ts x suc hthat B x hits X (ha vinganon-empt yin tersection). X B = f x : B x \ X 6 = ;g Bycom biningerosionanddilationoperations,t w onewoperators,openingand closing,canbecreated.Theopeningoperatorremo v esisolatedobjectsandbreaks w eakconnectionsincomponen ts.Itisanerosionoperationfollo w edb ydilation,using thesamestructuringelemen t:( X B ) B .Theclosingoperator,adilationfollo w ed b yanerosionusingthesamestructuringelemen t,( X B ) B ,connectssmallgaps incomponen ts.Inthisw ork,whenev eroneoftheseoperationsw asemplo y ed,the structureelemen tusedw asc hosentominimi zetheamoun tofc hangefromtheoriginal imagewhilestillac hievingthedesiredeect. Whenconsideringt w obinaryimages, X and Y ,their union ( [ )isthebinary imageinwhic heac hpixel x isamem berineither X Y ,orboth. X [ Y = f x : x 2 X OR x 2 Y g Inthissystem,theunionof X and Y iscalled\merging" X and Y The in tersection ( \ )of X and Y isthebinaryimageinwhic hallpixels x belongtoboth X and Y X \ Y = f x : x 2 X AND x 2 Y g 2.7.2 ImageReection Thegeometriccompleme n tofabinaryimageiscalleditsreection,producing amirrorimageofthebinaryimagerelativ etotheoriginoraxisofreection[55,48]. Giv enabinaryimage X ,itsreection B 0 issymmetri calwith B .Thatis: X 0 = f)Tj/T17 1 Tf64.0001 0 TD(x j x 2 X g

PAGE 43

28 1 2A3 4(a) 123 4A5 678(b) 2 c 1 a b (c) 2 c 1 a b (d) Figure7.ConnectedComponen tsinImageSpace. Bac kgroundpixelsarewhite, whileforegroundpixelsarenon-white.Figures(a)and(b)sho wthefourandeigh t-wiseneigh borsofapixelrespectiv ely .Figure(c)sho wsabinary imagewitht w ospatialregions,while(d)sho wstheircorrespondingeigh twiseconnectedcomponen tsmark edwithdieren tpatterns. 2.7.3 ConnectedComponen ts TheMRimagesprocessedherearearrangedasa256 256rectangulargridand apixelisconsideredtobe fouror eigh t-wise connected,ifithasthesameproperties (e.g.,classv alue)as one ofitsfouroreigh tneigh bors,assho wninFigures7(a)and (b)[54].Connectivit yisatransitiv erelation:ifpixels a and b areconnected,and b and c areconnected,then a and c arealsoconnected,regardlessofwhether a and c arethenearestfouroreigh tneigh borsofeac hother.Thet ypeofconnectivit ym ust beconsisten t.If a and b arefour-wiseconnected,then b and c m ustbefour-wise connectedalso. Figure7(c)sho wsanexamplebinarymask.Lookingatpixel a inRegion1,pixel a issaidtobeconnectedto a becausethereexistsapathfrom a to b thatiscomposed en tirelyofforegroundpixels.Pixel a cannotreac h c inRegion2ho w ev er,without crossingabac kgroundpixel.Thisconnectivit yofpixelsallo wsspatialcomponen tsto beformed,suc hasinFigure7(d),whic hindicatesthedistinctspatialregionsof(c) withdieren tpatterns. Connectedpixelsaregroupedtoformconnectedcomponen tsorregions,whic h areusedtoextractboundariesandindividualspatialregionsinanimage[54].Once

PAGE 44

29 spatialcomponen tsha v ebeencreated,eac hcomponen tcanbeisolatedandexamined, notonapixellev el,butonacollectiv elev el.Th usthesixpixelsinRegion2ofFigure7 canbeexaminedasagroup,independen toftheotherforegroundpixels.Inthisw ork, eac hspatialcomponen treceiv esaunique\label,"andisreferredtoas Region a =1 ;::: ;n where n isthetotaln um berofconnectedcomponen tsintheimage. Theconnectedcomponen tsoperationisalsousedb ythissystemtoll\holes" withinaforegroundmask.Smallholescannormallybelledwithasuitablysized dilationoperation.Largerdilationoperations,ho w ev er,candistorttheboundaries oftheoriginalbinaryimage,whileusingconnected-componen tsa v oidsthisproblem. Giv enabinaryimagemask,itispixel-wisein v ertedsothatan yforegroundpixel no woccupiesthebac kground,andvisa-v ersa.Aneigh t-wiseconnectedcomponen ts operationisperformedandthelargestforegroundregion(theairsurroundingthe head)isremo v ed.Allotherspatialcomponen ts,holesintheimagemask,aremerged withtheoriginalimagemask,andtheresultan timageinpixel-wisein v ertedagainso thatpixelsofin terestareforeground. 2.7.4 MedianFilter Giv enanimage,eac hpixeloftheimageisreplacedintheoutputimageb y themedianofthepixelscon tainedb yawindo woperator.At ypicalwindo wis3 3, butthissystemalsoused5 5and7 7windo ws. Medianltersareeectiv e inremo vingimpulsenoiseinimages[54]. Inthisdissertation,medianltersare appliedtobinaryimagestoremo v esmallisolatedareasforsimplifyingconnected componen tsoperationsandtosmoothanobject'sboundaryafteramorphological operation. Whilemedianlterscanoperateinagra yscaleen vironmen t,theyare usedhereinabinarycon text.

PAGE 45

30 Av ariationoftheerosionoperationthatmimic sthemedianlterw asalso added. Called\wiping,"itw asdesignedtoremo v esmallorisolatedpixels,while lea vingmostpixelsconnectedtolargercomponen tsin tact,withouttheneedforthe medianlter'ssortingstep.Giv enan m n windo w,withthecen terelemen tthe pixelunderconsideration,thetotaln um berofforegroundpixelswithinthewindo w iscoun tedandcomparedagainstathreshold. Ifthetotaln um berislessthanthe threshold,thecen terelemen tisremo v ed.Otherwise,itisleftinplace. 2.7.5 Thresholding Theimageprocessingtoolsusedabo v econsiderpixelsonlyinabinaryandspatialcon text,withoutregardtotheparticularfeaturev alueseac hpixelmigh tcon tain. Oneofthemostimportan tproblemsinimageprocessingistoseparateforeground regionsofin terestfromtherestoftheimageandthresholdingisoneofthesimplest \object-bac kgroundseparation"toolsa v ailable[55].Giv enagra yscaleimage,orin thecon textofm ultispectraldata,aparticularfeatureimage F ,abinaryimage F T canbecreatedwherethecorrespondingfeaturev alueofeac hpixelin F T isgreater thanathreshold T .F ormally: F T [ i;j ]= ( 1 if F [ i;j ] T 0 otherwise TheMRdomainisam ultispectraldomain,sothresholdingcanbedoneonan ysingle featureimage.Th us,apixelsurviv esonlyifitssignalin tensit yv alueinaparticular featureisgreaterthanthein tensit ythresholdforthatfeature.

PAGE 46

31 d BT = d w BT = w w1 BT = w1(x,y) (x,y) (x,y) l w2 Figure8.Bi-orthogonalThic kness.TheBT'softhreeexampleobjectsaresho wn. 2.7.6 Bi-orthogonalThic kness Bi-orthogonalthic kness, BT ( x;y ),w asdenedb yLi[70,71]tobethelesser ofthev ertical(Y)andhorizon tal(X)thic knessofaspatialobjectorcomponen tat poin t( x;y ).Inmostcases,( x;y )isthecen troidoftheobject.Figure8sho wsthe BTofsomeexampleobjects. TheBTofthecircularobjectisitsdiameter,while therectangularareahasaBTofthelesserofitswidthandlength.TheBTofthe irregularobjectisthelesserof w 1and w 2. T oapplybi-orthogonalthic knesstoa tissueorclusterwithm ultiplespatialcomponen ts,BTisdenedtobethemaxim um BT ( x;y )ofallcomponen tsintheimage.F orexample,ifthethreeobjectssho wnin Figure8w ereinthesameimageandbelongedtothesameclass,thentheBTofthat classw ouldbethemaxim umof d w and w1 .BT'sareusedtodetectwhitematter splittinginSection4.3.3.1. 2.8 SystemOv erview T obetterillustratethesystems'organizations,theyarepresen tedatamore abstractlev el.Thesystemsforbothpathologydetectionandtumorsegmen tation emplo yaparadigmcalled\iterativ eprocessing"whereaparticulargoalisac hiev ed

PAGE 47

32 throughaseriesofprocessingstages,ratherthanasingleclassicationstep.Eac h oftheseprocessingstageslocatesandremo v eseasilyiden tiablepixels,allo wingthe systemtoconcen trateontheremaining(few er)pixels. 2.8.1 P athologyDetectioninSlicesBelo wtheV en tricles Figure9sho wstheprimaryprocessingstagesforpathologydetectioninthe \lo w er"slices.Chapter4giv esamoredetaileddescriptionoftheseprocessingsteps. Afterclusteringthera wdatawiththeF CMalgorithm,theinitialsegmen tation ispassedtoStageOne,whic hseparatesclustersprimarilycon tainingpixelsbelonging toextra-cranialtissuesfromthoseclusterswithmostlyin tra-cranialpixels. This separationisbasedonanatomicalkno wledgeconcerningthespatialorganizationof extra-cranialtissuessuc hasskin,fat,andm usclesv ersusparenc h ymalbraintissues andCSF.Aninitialin tra-cranialmaskiscreatedandrenedb yremo vingextracranialpixelsthatw eremisplacedin toanin tra-cranialcluster. StageTw odeterminestheT emplateoftheslice,basedinitiallyontheshapeof thein tra-cranialmask,thenthroughthein ternalstructuresofthebrain'ssofttissues. OnceitsT emplatehasbeendetermined,thecorrespondingqualitativ emodelcanbe assigned. InStageThree,eac hpixelwithinthein tra-cranialmaskismappedwithitsclass labelfromtheinitialF CMsegmen tationstep,allo wingeac hin tra-cranialclusterto beanalyzed,butonlywiththosepixelstrulybelongingtothein tra-cranialregion. Thesystemthencomparestheslicewithitsqualitativ emodel.Ifadeformationis found,thesliceislabeledasabnormalandsen ttothetumorsegmen tationsystem. Otherwise,thesliceislabelednormalandprocessinghalts.

PAGE 48

33Raw MR image data: T1, PD, and T2-weighted images. Cluster labels are mapped back onto pixels within the intra-cranial mask. Each cluster can then be examined separately. Deviations from qualitiative models denote abnormal slice (such as the failure of symmetry along the vertical axis, shown below). Abnormal slices are passed to the tumor segmentation system. STAGE THREE The slice's template is initially identified based on the shape of the intra-cranial mask. The mask is refined to verify expected intra-cranial tissues, which are then used to confirm the slice's template. STAGE TWO Separation of clusters containing primarily extra-cranial tissues from clustering with mostly intra-cranial tissues forms an initial intra-cranial mask. The brain is extracted and the mask refind based on spatial knowledge. STAGE ONE Initial segmentation by unsupervised clustering algorithm. Mask after initial cluster separation. Extracted intracranial mask. Pixels mapped onto intracranial mask. Detected abnormality.Figure9.Ov erviewofP athologyDetectionforSlicesBelo wtheV en tricles.

PAGE 49

34 2.8.2 P athologyDetectioninSlicesAbo v etheV en tricles Figure10sho wstheprimaryprocessingstagesforpathologydetectioninthe \upper"slices.Allofthekno wledgedescribedinSection2.2appliestobothupperand lo w erslices,asdoessomeanatomicalkno wledgedescribedinSection2.3.Therefore, man yoftheprocessingstepsdescribedinChapter4forprocessingslicesbelo wthe v en triclesareapplicablehereasw ell.Amoredetaileddescriptionofthesepathology detectionstepsb ythisauthorcanbefoundin[20,18]. 1.Afterclusteringthera wdatawiththeF CMalgorithm,clustersprimarilycontainingpixelsbelongingtoextra-cranialtissuesareseparatedfromthoseclusters withmostlyin tra-cranialpixelsusinga\quadrangletest,"similartotheone describedinSection4.2.1. 2.Oncein tra-cranialclustersareiden tied,whitematterclustersarelocatedusing kno wledgethatwhitematterhasthelo w estT2v alueofallin tra-cranialtissues (Section4.2.3).P ossiblewhitemattersplittingisdeterminedwithat w o-lev el binarydecisiontreerstdev elopedb yLiin[70,71]andusedinSection4.3.3.1 whendetectingatransitionbet w eenT emplate1and4slices. 3.Iftheslicebeingprocessedisnottheinitial/cen terslice,CSFislocatedb y searc hingforthein tra-cranialclusterwhosecen troidisclosestinT2spaceto thecen troidoftheCSFclusterintheinitialslice(whic hhasbeencompletely processed).P athologycanalsobedetectedb ynotingthatanormalCSFcluster shouldha v ealo wT1v alueforitscen troid,whileaclustercon tainingenhancing pathologywillha v eahighT1v alueforitscen troid. 4.WhitematterandCSFareisolatedandaslicetemplateisassignedb ycoun ting then um berofsignican tregionsofCSFand\holes"inwhitematter(regions ofCSFsurroundedb ywhitematter),determinedb ywheretheslicein tersects withthev en tricles.F orexample,aT emplate3sliceliesabo v ethev en tricles andhasnoholesinwhitematter.Iftheinitialsliceisbeingprocessed,itcan beassumedthesliceisT emplate1. 5.Onceatemplateisassignedtotheslice,pathologydetectionbeginsb yisolating whitematterandusing\emitterlines"(Sections4.3.3.2and4.4.3.2)tov erifythe shapeofwhitematteraroundtheV Aandalongthein ter-hemispheri cssure, whic hisusuallydeformedinthepresenceofgrosspathology .Thesymmetryof whitematteralongthev erticalaxisisalsov eried.

PAGE 50

35 Raw MR image data: T1, PD, and T2-weighted images.White matter is isolated and refined to examine its shape. Deviations from qualitiative models denote an abnormal slice. If processing the initial slice and white matter is normal, CSF (above) is located by approximating the ventricle area and finding the cluster with the most pixels inside. Both CSF and white matter should be vertically symmetrical. Abnormal slices are passed to the tumor segmentation system. The white matter cluster(s) (left) are located and the slice's template is assigned based on the number of significant CSF regions (right) and "holes" in white matter caused by CSF regions. If processing the initial slice, the slice is skipped and assumed to be Template 1. Separation of clusters containing primarily extra-cranial tissues from clustering with mostly intra-cranial tissues forms an initial intra-cranial mask. The brain is extracted and the mask refined based upon spatial knowledge. Initial segmentation by unsupervised clustering algorithm. Image After Removing Extra-Cranial Clusters The initial slice is skipped in this step also because CSF has yet to be located. CSF in the initial slice is located below and used as a landmark in finding CSF in other slices. White matter here is deformed by the presence of pathology. Each gray matter cluster is isolated and examined with a bi-orthogonal thickness measure. Normal gray matter should have no regions that are spatially "dense." Abnormal slices are passed to the tumor segmentation system. The image on the right contains pathology.Figure10.Ov erviewofP athologyDetectionforSlicesAbo v etheV en tricles.

PAGE 51

36 6.Iftheinitialsliceisbeingprocessed,thev en tricleareaisappro ximatedwith \c haincoding"[54]thatsearc hesalongthewhitemattersurroundingtheV Afor thelocalextremaofeac hofthefourhornsinthev en tricles.Figure23(a)sho ws anexampleofthisappro ximation.CSFislocatedb yndingtheclusterwith thegreatestn um berofpixelswithintheappro ximatedV A.Inalltemplates, theCSFclustershouldalsobethein tra-cranialclusterwiththehighestT2w eigh tedcen troidandshouldalsobespatiallysymmetricalongthev erticalaxis. Violationsofeitherconditionindicatepathology 7.Gra ymatterclustersarelocatedbet w eenwhitematterandCSFinT2space. Eac hgra ymatterclusterisisolatedandtestedforpathologyb ynotingthat anormalgra ymatterclusterisrelativ ely\sparse"spatially ,asdenedb ybiorthogonalthic knessmeasure(Section2.7.6),whilepathologyismore\compact"groupedinimagespace. 2.8.3 T umorSegmen tation Figure11sho wstheprimarystepsthattak eplaceinextractingtumorfrom ra wMRdata.Chapter5describestheindividualprocesssteps.Allslicesprocessed herehadabnormalitiesdetectedwithinthemandarepassedon w ardtothetumor segmen tationsystemalongwithkno wledgegainedduringpathologydetection.Slices thatarefreeofabnormalitiesarenotprocessedfurther.Figure11referstopathology detectionasStageZero. Thetumorsegmen tationsystemhasv eprimarysteps. StageOnecreates animagemaskofthein tra-cranialregionfromtherestoftheMRimagebasedon informationpro videdb yStage0andrecapturesan ylostpathologicalpixelsmisclassiedin toanextra-cranialclusterduringinitialsegmen tation.Sincepixelsofair,fat, m uscle,etc.arenotofin terest,onlypixelswithinthismaskareconsideredinStage Tw o. Aninitialtumorsegmen tationisproducedinStageTw othroughacom bination oftheapplicationofhistogramthresholdsontheT1andPDfeatureimages.The actualthresholdsusedarenotxed,butautomaticallyadjusttotheslice'sproperties. Theinitialtumorsegmen tationispassedontoStageThree,whic hprocessesonly

PAGE 52

37Raw MR image data: T1, PD, and T2-weighted images. Radiologist's hand labeled ground truth tumor. Tumor segmentation refined using ``density screening.'' STAGE THREE Initial tumor segmentation using adaptive histogram thresholds on intracranial mask. STAGE TWO Intracranial mask created from initial segmentation to reclaim possible lost tumor pixels. STAGE ONE STAGE 0 Pathology Detection. Slice tissues are located and tested. Slices with detected abnormalities (such as in the white matter class shown) are segmented for tumor. Slices without abnormalities are not processed further. Initial segmentation by unsupervised clustering algorithm. White matter class. Removal of ``spatial'' regions that do not contain tumor. STAGE FOUR Final thresholding in T1 spectrum. Remaining pixels are labeled tumor and processing halts. STAGE FIVEFigure11.Ov erviewoftheT umorSegmen tationSystem.

PAGE 53

38 thosepixelscon tainedinthenewmask.A\densit yscreening"operationinStage Threeremo v esmoreun w an tedpixelsbasedontheobserv ationthatpixelsofnormal tissuesaregroupedmorecloselytogetherinfeaturespacethantumorpixels. StageF ourcon tin uestumorsegmen tationb yseparatelyanalyzingeac hspatially disjoin t\region"inimagespacecreatedb yaconnectedcomponen tsoperation.Those regionsfoundtocon tainnotumorareremo v edandtheremainingregionsarepassed toStageFiv eforapplicationofanalthresholdintheT1spectrum.Theresulting imageisconsideredthenaltumorsegmen tationandcanbecomparedwithaground truthimage.

PAGE 54

39 CHAPTER3 RELA TEDW ORK Theextensiv euseofmagneticresonancebrainimaginghasdra wnincreasing atten tiontothesubjectofthesegmen tationandlabelingofMRslicesandv olumes. Th us,av ariet yofapproac hestotheproblemofsegmen tingMRbraindataha v ebeen tak en,thoughtheseeortsremainlargelyexperimen tal,especiallyinpathological cases.Thesegmen tationtec hniquesthatha v ebeenexploredcanbeplacedin tot w o primarycategoriesbasedonthekindofdatatowhic htheyareapplied:singlecon trast data,whereonlyasingleMRpulsesequencew asused,andm ultispectraldatawhere morethanoneMRpulsesequenceisused.Thisc hapterwillbrieyreviewsomeof thoseeorts. Man yoftheseeortsdealonlywithnormalcases.Theprimaryfocusofthis dissertationissegmen tationofenhancingtumor,makingcomparisonsbet w eenthe t w oinappropriate.Thiskno wledge-basedsystemcandocompletetissuelabelingof someslices,ho w ev er,resultsofwhic hcanbefoundb ythisauthorin[20,18]. Of thesegmen tationeortsthatdealwithpatien tcases,mostdonotquan tifytheir performanceagainstgroundtruth,insteadusingothermethodssuc hasvisualinspectionor\agreeabilit y"bet w eensegmen tationso v erm ultipletrials.Thoseeorts withquan tiableresultsarereview edhere,thencomparedinSection6.3.4withthe kno wledge-basedsystemusingtheev aluationform ulaoftheeortreview ed. Itshouldalsobenotedthatsomeoftheeortsdealingwithpatien tv olumes segmen tm ultiple -sclerosis(MS)lesions. SinceMSlesionsarenotcancerous,but

PAGE 55

40 regionsofwhitematterdem y el iniz ation,theyha v edieren tproperties(suc hasa lac kofsignican tedema,whic hcanblurtumorboundaries)thangliomas.Therefore, comparabilit ybet w eensuc hsystemsandthekno wledge-basedsystempresen tedhere islimitedsince,ineect,theyareperformingdieren ttasks. 3.1 SingleCon trastSegmen tationMethods Methodsha v ebeendev elopedforMRsegmen tationwhereonlyasinglepulse sequenceisa v ailable.Inthesecases,theMRimageistreatedasagra yscaleimageand canbeprocessedwithan um berofimageprocessingtools.Unfortunately ,whilethese methodscanpro videusefulinformationandcansegmen tv arioustissuet ypes,they areusuallylimitedtotissueswithrelativ elysimplestructures[36]. Morecomplex tissuet ypes,suc haspathology ,generallyrequireadditionalinformationthatisfound inm ultispectralMRdata[23]. Approac hesthatusem ultispectraldata,ho w ev er, suc hasthesystempresen tedhere,canusethesesinglecon trasttec hniquesina m ultispectralcon textb yapplyingthemonm ultiplefeatureimagesandcom bining theirresults. 3.1.1 Thresholding Tsai,Majunath,andJagadeesan[113]presen tanunsupervisedsystemthat detectsm ultiple-scl erosis(MS)lesionsinPDandT2-w eigh tedimages.Kno wledge intheformofanatomical/structuralinformationisusedtoguidelo wlev elimage processingoperations,suc hasmorphologicaloperations,toextractthein tra-cranial region.Histogramanalysisisalsoincludedtoallo wkno wledge-basedthresholdingof braintissues.Ov er200images,mostlyintheaxialplane,w eretak enfromeigh tcase studies(t w onormal,fourwithabnormalv en triclesizes,onewithm ultiplelesions),

PAGE 56

41 butnocleardistinctionbet w eentrainingandtestsetsw asmade.Thresholdselections w erebasedonhistogramanalysis,bothinclinicalstudiesandempiricalobserv ations, ofbraintissues.Whiletheauthorssho wsuccessinlabelingnormalsanddetectingthe MSlesions,theresultsarenotdirectlycomparedagainstan ygroundtruth.Instead, theystatethesegmen tationsareof\acceptablequalit y"accordingtoradiologists. Moreo v er,theauthorsonlyconsiderslicesthatin tersectwiththev en triclearea. In[107],SukukiandT orik a w ause\iterativ ethresholding"toextractthein tracranialregioninaxialslices.Tw oimagesetsw erea v ailable.Therstconsistedof fourteenslicesacquiredusingIn v ersionReco v ery ,whilethesecondsetconsistedof 128T1-w eigh tedslices.A\goodnessmeasure"c hec ksanatomicalboundariesofthe resultingsegmen tationandadjuststhethresholdsifimpro v em en tisneeded. The systemac hiev esthehighestgoodnessmeasureinsev enofthefourteenslicesfrom therstset,andsev enofthe128slicesinthesecondset.Theremainingslicesfrom therstsetand115slicesfromthesecondac hiev edasub-optimal,butsatisfactory segmen tation.SinceonlynormalMRslicesaretested,itisunkno wnifthemethod isapplicabletopathologicaltissues. AlthoughnotinMRimagesofthebrain,similaruseofkno wledgeinthresholdingw asdoneb yKobashiandShapiro[64,63]inextractingabdominalorgansfrom CTimagesonatrainingsetof100imagesfromv epatien tsandtestsetof75images fromthreeadditionalpatien ts.Segmen tationw asgradedqualitativ elyb ytheauthor usinggradesA,B,orCtoindicatethesev erit yofmismatc hes.Kidneysegmen tationsw eregiv enanAin94%oftheslices,whilethespleenandliv ersegmen tations w eregiv enanAin62%and60%oftheslicesrespectiv ely .Noquan titativ ev alues w eregiv entothegradesb ytheauthors,ho w ev er,sotheexactsegmen tationqualit y requiredforanAisunkno wn. GongandKulik o wski[42]comparelocalbinarythresholdsv ersusglobalm ultilev elthresholdsforthetasksofextractingthein tra-cranialregionfromanMRimage

PAGE 57

42 andsegmen tingtumorsandMSlesions.F ort yPD-w eigh tedandfort yT2-w eigh ted imagesfromtenpatien ts(fourslicesfromasequenceofscansfromeac hpatien t)w ere usedtotestthethresholdmethodsforextractingthein tra-cranialregionfromanMR image.F ort y-onePDandtheirrespectiv eT2-w eigh tedimagesfromv epatien tsw ere usedtotesttumorsegmen tation,whileonlyeigh tT2-w eigh tedimagesw erea v ailable forsegmen tingMSlesions. Theauthorsstatethethethresholdingmethodsw ere dev elopedinotherw orks,anddonotindicateifan yoftheimagestestedherew ere usedastraining.Althoughthedataw asm ultispectral,thethresholdingoperations w ereappliedtothePDandT2imagesseparatelyandthem ultispectralinformation w asnotexploited. Accordingtotheauthors,thelocalthresholdingmethodw as superiortotheglobalmethodinsegmen tingMSlesions.Thelocalmethod,ho w ev er, performedrelativ elypoorlywhensegmen tingbraintumors. Theauthorscitethe localmethod'sdependenceonedgedetectabilit yandwithinsliceinhomogeneities andnotetheglobalthresholdingmethodassuperior.Basedonwhetherthetarget regionw asseparatedfromsurroundingstructureswithouto v er-segmen tation,high performanceratesaresho wn.Itshouldbenoted,ho w ev er,thattheirmethoddidnot remo v eenhancingextra-cranialtissuesorotherin tra-cranialregionsspatiallydisjoin t fromthetumormass,bothofwhic hareaddressedb ythekno wledgebasedsystem. F urthermore,nodirectcomparisonwithan ygroundtruthw asmade,sotheirsystem's segmen tationaccuracyisunkno wn. Brummer[12]proposesasystemthatrecomputesthresholdv aluesfrominitial sub-optimalsegmen tationsusingparametriccurv ettingandhistogrammodicationtoestimatetruetissuedistributions. Thisapproac hw asappliedtocoronal T1-w eigh tedimagesofsixnormalbrainv olumes,appro ximately35-40slicesperv olume.Itw asalsoappliedtoanarticialphan tomofthev en tricles.F orthev en tricle phan tom,theresultssho wthesystem'sabilit ytoproperlyrecomputetheoptimal

PAGE 58

43 thresholdforextractingthearticialv en triclesgiv enaninitialsub-optimalsegmentation.F orthenormalv olumes,theauthorsho wsho ww ellthecurv ettingprocedure correspondstotheactualT1-w eigh tedhistogram.Onlyoneofthesixcasessho w ed poorcorrespondencebet w eenthettedcurv eandtheactualhistogram.Noquantitativ ev alidationw asmade,nordidtheauthorindicatewhattrainingw asusedor ho ww ellitw ouldperformincaseswithgadoliniumenhancemen t. P annizzo,Stallmey er, etal. usehistogramanalysistoseparateparenc h ymal tissuesfromperiv en triculareusions,c hronicMSlesions,andedema[94]. Before thisanalysisisapplied,ho w ev er,a supervised edge-follo wingalgorithmisapplied toextractthein tra-cranialregion. Also, operatorin terv en tion isalsoneededto determinewhethertheslicehasam ulti-exponen tialhistogramdistribution.Atotal ofsev en t y-t w oslicesfromtenpatien tsw ereused,withallpatien tsha vingatleastone follo w-upscanforatotaloffourteentransitionsbet w eenscans.MultispectralMR scansw eretak en,butthemethodappearstouseonlyasinglecon trastimage.Tw o independen toperatorssho w edameanagreemen tof97%withastandarddeviation of2 : 9%.F orv alidatingthegro wth/shrink ageoftheMSlesionso v ertime,themean resultsofthesystemw erecomparedwithclinicalstudiesandagreemen tw asfound int w elv eofthefourteentransitions.Noattemptw asmade,ho w ev er,todistinguish bet w eennewMSplaquesandoldMSplaques,edema,orotherlesions. LimandPerbaumfocusonmanipulatingT1andT2images,b yapproac hes suc hasimagedierencing,toenhancetissueseparationbeforein teractiv elyapplying thresholdsforsegmen tation[74].Thirt y-fourslicesfromv enormalmalesubjects w ereexaminedandresultsw erebasedonaratioofthen um berofgra ymatterpixels towhitematterpixels.Theseratiosw erethencomparedwithsimilarratiostak en frompost-mortemstudies.Allratiosgeneratedb ytheauthors'system(1 : 07-1 : 18) w erefoundtobewithintherangeofv aluesgiv enb ythepost-mortemstudies(0 : 9

PAGE 59

44 -1 : 3).T raining/testslicedistinctionw asnotmadeandnopathologicalcasesw ere ev aluatedandonlynormalv olumesw ereprocessed. Finally ,Kundu[66]tak esanalternativ eapproac htohistogramanalysisand localthresholdingb yusingan\optimalquan tizer"designtosetthresholdsinCT scanandx-ra yimageswithoutha vingtocomputetheimagehistograms. Kundu statedthatthenewdesignw ascomputationallylessexpensiv ey etproducedresults comparabletostandardhistogrambasedmethods. Theauthor,ho w ev er,didnot pro videthen um berofimagesthismethodw asdev elopedortestedo v er. 3.1.2 T extureAnalysis Kjr,Ring,Thomsen,andHenriksen[62]look edforsignican tdierencesbet w eentissuet ypes,bothnormalandpathological,in6v olun teersand88patien tcases usingav ariet yoftexturemeasuremen ts.Supervisedregionsofin terest(R OI's)of eac hbraintissuew ereselectedfroman um berofparameterimages,includingcalculatedT1andT2images(imagesthatappro ximatethe\pure"T1andT2properties ofbraintissues,andarecalculatedfromT1,PD,andT2-w eigh tedimages),and\amplitude"protondensit yimages.Aseriesofrstandsecondordertexturemethods w erethenappliedtoeac hoftheseR OI's.Theauthors'goalw asasimplestudyofho w w elleac htexturemethodw oulddiscriminatebet w eent w otissuet ypes(e.g.,white mattervs. CSF)anddidnotev aluatethemonho ww elltheyw ouldsegmen tthe en tireslice. TheF CMalgorithmw asmodiedb ydeOliv eiraandKitney[29]tousetexturebasedfeaturesinsteadofsignalin tensit yv aluesinanunsupervisednonparametric en vironmen t.Sixt y-foursagittalslicesofah umanheadw ereprocessed,butnoinformationconcerningtheirimagingparametersw aspro videdb ytheauthors.F rom thesesixt yfoursagittalslices,asingletransaxialslicew asgeneratedandatotalof

PAGE 60

45 fourtexture-basedfeaturemeasuremen tsw erethencalculatedandaddedtotheoriginalimage,resultinginav e-dimensionaldataset.Beforetexturesw ereextracted, ho w ev er,theslicerstunderw en taseriesofsmoothingoperationstoenhancetexture uniformit ywhilestilla v oidingsignican tblurring.Theauthorsbasedthequalit yof resultsuponvisualinspectionanddidnotcomparethemwithan ygroundtruth. Zuna,H arle,Sc had, etal. [101]useda\la y ered"approac htoiterativ elysegmen t specictissuesinthebrainimageb yusingdieren ttexturemeasuremen tsindieren tfeatureimages.Tw elv epatien tswithbraintumorsw ereexamined;withkno wn tumorhistologiesa v ailableintenofthesepatien ts. Fiv epatien tsw erediagnosed withprimaryglioblastomaswhile,v eothershadtumorsasaresultofmetastases. Atotalof113texturesamplesw ereselectedb yradiologistsfromcalculatedT1and T2parameterimages,buttheauthorsstatethateectiv eR OI'susuallyrequireat least200pixels,whic hreducedthen um berofsamplesactuallyusedto78. These tissuesamplesw erethendiscriminatedinahierarc halfashion,whereeac h\la y er" addressesaparticulartissuet ypeandusesadieren ttexturefeatureinitsdecision makingprocess. Theauthorssho wdiscriminationratesof95%orbetterbet w een theR OI'sselected,butdonotconsideran ytesttissuesamples.Sincetextureisa statisticalfeaturethatrequiresalargen um berofpixelsforareliablev alue[46],itis bettersuitedtoregioniden ticationandlabelingthanboundarydelineation. Mark o vrandomeldsha v ebeenstudiedextensiv elyasamodelfortexture represen tation[55].\Mixels"w eredenedb yChoi,Ha ynor,andKim[17]asv o xels (pixelswithv olume)thatsueredfromthepartialv olumeeect,whenm ultipletissue t ypesarecon tainedinthesamepixel/v o xel.ByassumingaGaussiandistributionin signalin tensit y ,aMark o vrandomeldisusedtocreatea\mixel"imagemodelto statisticallyclassifythemajorin tra-cranialtissuesinpartialnormalbrainv olumes, basedonman uallyselectedtrainingpixels. Theauthorsdonotstatethen um ber oftrainingimagesusedandtestingw asperformedb ycreatingsim ulatedbrainMR

PAGE 61

46 imagesbasedonthemeantissuedistributionsofthetrainingcaseswithwhiteGaussiannoiseadded.UseofMark o vrandomeldsha v ealsobeenattemptedb yLeah y Hebert,andLee[68].Some,ho w ev er,ha v enotedthataccurateestimationofmodel parametersisoftendicult[129].A ttemptingtoestimateparametersforstructures withfaralessuniformdistributionthannormalin tra-cranialtissuesw ouldbeev en moreproblematic. 3.1.3 EdgeDetection Bomans,H ohne,Tiede,andRiemer[8]extendaMarr-Hildrethoperatorin to 3Dandsho wthattheoperator'szero-crossingsarerelatedtoanatomicalsurfaces. Fiv eimagev olumeswith128slicesperv olumew eretested,withthesignalin tensit y v alueslinearlyscaledfrom12to8bitsandin terpolatedtoac hiev eisotropicgra y lev elv alues.Theseimagesw erealsoexaminedtodeterminetheoptimalsizeofthe Marr-Hildrethoperator.Morphologicalltersandaconnected-componen tsalgorithm arethenusedtorenethedetectededges.Con tourcorrectionandlabelingisperformedin teractiv ely ,sonodirectgroundtruthcomparisonw asmadeandonlynormal v olumesw ereprocessed. W uandLeah y[131,130]usenet w orko wtheorytoimpleme n tedgecon tour ndingthato v ercomestheshortcomingsofMarr-Hildrethorlocalin tensit ydierence models.Ahierarc haladjacencygraphisconstructedfrompoten tialedgeelemen ts andeac hv ertexrepresen tsahomogeneousregionintheimage.Net w orko wisused tondgrapharcswiththegreateststrengths(probabilit ythatt w ov erticesbelong tothesametissue)andremo v earcswithw eak erstrengths.T umorsegmen tationis sho wn,butonlyonePD-w eigh tedslicew asprocessed,sothemethod'srobustness o v erarangeofimagesisnotkno wn.

PAGE 62

47 Gibbs,Buc kley ,Blac kband,andHorsman[41]examineT1-w eigh tedimages andsegmen tenhancingin ter-cerebralgliomasinsev en teenv olumesoftenpatien ts. Onlysliceskno wntocon taintumorareconsidered.Supervisedthresholdingappro ximatingtheboundarybet w eentumorandnon-tumortissues,follo w edb yaSobel lter,pro videstheinitialsegmen tation.TheedgestrengthoftheSobellteristhen addedasinformationforaregiongro wingalgorithmtorenethetumorboundaries. Statisticalmeasuremen tsandanearestneigh borltercompletethesegmen tation. Experimen tsw ererunwitht w ooperators,withoneandthreetrialsrespectiv ely .T umorv olumesarecalculatedforsegmen tationsusingonlythethresholdmethodand aftertheregiongro wingsteprespectiv ely ,tocomparetheirlev elofagreemen t.Of thesev en teenv olumesstudied,onlyonecasew assho wntoha v e\clear"disagreemen t bet w eenv olumeswithandwithouttheregion-gro wingstep,whic hw asattributedto inferiorenhancemen toftheimage.Noclearbenc hmarkfor\agreemen t"ispro vided, ho w ev er.Threepatien tshadfollo wupscans,t w o,one,andonerespectiv ely .Ofthe fourtransitions,onlyonetransitionw asuniformlytrac k ed.Withoutground-truth, ho w ev er,itisdiculttoquan titativ elyev aluatetheirsystem'sperformance. Dellepiane[30]andRaandNewman[96]suggestthatedgedetectionmethods areunlik elytopro videreliablesegmen tationofcomplexstructureslik etumors.This w ouldholdespeciallytruefortumorswithextremelydiuseboundaries.Infact,Gong andKulik o wski[42]citeedgequalit yastheprimaryreasonforthepoorperformance oftheirlocalthresholdingmethod. 3.2 MultispectralSegmen tationMethods Multispectraldatasetscon tainaseriesofdieren tpulsesequenceimages,eac h withitso wntissuecon trastc haracteristics.Therefore,segmen tationmethodsshould beabletoexploitthem ultidim ensionalinformationcon tainedb yeac hpixel.

PAGE 63

48 3.2.1 ClusteringandNonparametricP atternRecognition Av ariet yofclusteringtec hniquesha v ebeenin v estigatedindetailb yBezdek, Clark e,Hall,V elth uizen,V aidy anathan,Bensaid, etal. inan um berofpapers[6 25,26,126,23,114,117].Thesetec hniquesincludedbothhardandfuzzyc-means (F CM)[7],appro ximateF CM[13],iterativ eleastsquaresclustering[47],andsplit F CM[126].A\semi-supervised"F CM(ssF CM)[4]w asalsoexamined.ThessF CM algorithmallo wstrainingpixelstobeselectedasclassexemplarsand\w eigh ted"toinuencetheclusteringprocess.Acommonlyusedsupervisednonparametricclassier, thek-nearestneigh bors(kNN)algorithm,andacommerci allya v ailablesupervised seedgro wingmethod,ISGAllegro(fromISGT ec hnologies,T oron to,Canada),w ere alsoincludedforcomparisonpurposes.Resultingsegmen tationsw erecomparedwith radiologistlabeledgroundtruth. F romtheseeorts,an um berofconclusionsw eremade.Therstisthatsupervisedmethodsin troducethesignican tproblemofoperatorv ariabilit yinselecting trainingsets.Inestimatingtumorv olumes,thekNNandssF CMmethodssho w ed appro ximately9%and6%in tra-operatorv ariabilit y ,respectiv ely .In ter-operatorv ariabilit yw as5%forkNNand4%forssF CM.TheISGmethodsho w edanev enlarger v ariabilit y ,whic hw asashighas15%[117].Theotherproblemwithsupervisedmethodsisthatselectinggoodtrainingexamplescanbetimeconsuming,dependingon thecomplexit yofthedata. Sinceunsupervisedmethodsdonotusetraining,they donotsuerfromtheseproblemsandoer\reproducibilit y ,"meaningthatgiv en aparticulardataset,thesamesegmen tationwillbeac hiev edev erytimethedata isprocessed.Unsupervisedmethods,ho w ev er,canha v eproblemsarrivingatmeaningfulsegmen tations[126,123]. \V alidit y-guided"clustering(V GC)w asproposed in[3,123]toaddressthisproblemb yev aluatingthe\v alidit y"ofclusterstobemeasuredbasedonwithin-classhomogeneit yandbet w een-classseparation.WhileV GC

PAGE 64

49 didsho wimpro v em en to v erF CM,sometumorsw erestillundetected[123].Results ofthekno wledge-basedsystemarecomparedwiththosegeneratedb ykNN,ssF CM, andISGinChapter6. Thereviewsaddressedbothsegmen tationwithinasinglesliceasw ellastumor v olumeestimation.Additionaleortsw eremadeb yV elth uizen,Ph uphanic h, etal. in[126]tosegmen tthev olumeasanen tiredatasetusingkNNandbothhardand fuzzyc-meansclustering.Resultssho w ed,ho w ev er,thatthereappearstobeenough datain ter-sliceinhomogeneit ytoprev en tsatisfactorysegmen tation.Hall,Bensaid, et al. [45]comparedstandardandappro ximatefuzzyc-meansclusteringwithasupervisedfeed-forw ardcascadecorrelationneuralnet w ork.Theresultingsegmen tations w eresubmittedtoradiologistsforev aluationandwhiletherew assomein tra-and in ter-expertv ariabilit y ,thefuzzyc-meanssegmen tationsw eregenerallyconsidered superiortotheneuralnet w ork.Li[70,71]usedakno wledge-basedexpertsystemto automaticallydetectpathologyandlabelnormaltissues(afterF CMsegmen tation)in asmallrangeofMRslicesthatin tersectedthev en tricles.An um berofcon tributions madeinLi'sw orkareusedinthesystempresen tedhere. Problemswithobserv erv ariabilit yw erealsosho wnb yVinitskiand etal. in[127].ThekNNmethodw asappliedtoarticialphan tomsandh umanMRv olumes(v enormals,threepatien tswithtumor,fourpatien tswithMSlesions).In terobserv erv ariabilit yrangedbet w een6.6%and9.3%,whilein tra-observ erv ariabilit y w asfoundbet w een8.3%and10.4%.Kikinis,Shen ton,Gerig, etal. [60]usedman ual labelingandthekNNmethodtosegmen t15malev olun teers,onepatien twithMS lesions,andaneuro-surgicalpatien t. Fiv eobserv ersc hosethekNNtrainingsamplesandalsoman uallylabeledcertainslicesforcomparison.Man ualsegmen tations sho w edpixel-b y-pixelo v erlapof86 : 2%,whilethekNNsegmen tationsw erebetter with92 : 4%.Theauthorsalsosho w edthatin ter-observ erv ariabilit ywhenselecting trainingR OI'sgenerallydecreasesasobserv ersgainmoreexperience.

PAGE 65

50 Eortsin toclusteringbasedsegmen tationha v ealsobeenmadeb yotherresearc hers. \Con textualfuzzyclustering"(CTF CM)w asproposedb yKeh t,Ch un, Ha yman,andW endt[59].Kno wledgeisdirectlyin troducedin totheclusteringprocessb yadaptiv elyadjustingthemem bershi pmatrixofanindividualv o xelb yconsideringthemem bershipsofitssurroundingneigh bors.Themethodrequirestraining toestablishthekno wledgeconstrain ts,butthealgorithmisotherwiseunsupervised. Themethodw asrstappliedtoanimagephan tomcreatedb ysuspendingahardboiledc hic k enegginabo wlofgelatin.TheCTF CMsegmen tationsw erecompared totheresultsofotherclassiers,suc hasBa y esian,hardc-means,kNN,F CM,and thresholdedF CM.Theauthorssho wthattheCTF CMmethodbestestimatedthe v olume(basedonw aterdisplacemen t)oftheen tireegg,asw ellasy olkandeggwhite separately T ensetsofaxialMRslicesofthebrainw eretak enfromfournormal subjectsandthenclustered.Theresultingsegmen tationsw erethengiv entoaradiologist,whogradedeac hsliceonascaleof1to10.TheCTF CMmethodw assho wn toha v eana v erageratingof10.0,whilethesecondbestmethodw asF CMwithan a v erageratingof6.5.Althoughtheauthorsdonotstatethen um berofclustersused intheirexperimen ts,Section4.1discussestheproblemsencoun teredwithF CMwhen usingaone-to-onecorrespondencebet w eenthen um berofclustersandtissueclasses, asw ellasho wtheproblemw asaddressedinthissystem. Brandt,Bohan,Kramer, etal. [10]examinedh ydrocephalic(excessCSF,reducedwhitematter)c hildrenwithav arian toftheF CMalgorithmsuitedfortexture iden tication. OnlythePDandT2imagesw ereusedandthe m parametermentionedinSection2.5w asoptimizedfortheimagesetstheyused.Theauthorsalso man uallymask edoutallextra-cranialtissuesbeforeclustering,andremo v edlo w-lev el in tensit ynoiseusingoperatorthresholdingontheT2image.Resultsw erev eried b ycomparingthemeandistributionofbraintissuesofthreeh ydrocephalicv olumes

PAGE 66

51 againstthree\con trol"v olumesandsho w edthattheh ydrocephaliccasesdidha v ea reducedlev elofwhitematterandincreasedamoun tofCSF. Althoughtheydonotsegmen tMRimages,SridharandMurt y[105]alsointroducekno wledgein toclusteringwiththeir\deductiv eclustering"algorithm.An inheritancenet w orkisusedtoexplicitlystoreandrepresen tdomainkno wledgeabout inputobjects.Thisdomainkno wledge,bothconceptualandcon textual,guidesthe clusteringalgorithm,allo wingforh uman-orien tedsegmen tations. 3.2.2 Kno wledge-BasedSegmen tation Ra y a[97]designedarule-basedsystemtosegmen tandquan titativ elyestimate majorbrainstructures.Lo w-lev elfeaturesfromPDandT2imagesw ereextractedto enhancetheseparationofv o xels.Classesrepresen tingspecictissuesarethencreated witheac hv o xelintheimagegiv ena\condence,"similartoafuzzymem bership, ineac hclass.Thecondencelev elsestablishahierarc halorder,allo wingv o xelswith highercondencetoaidclassicationofv o xelswithlo w ercondencelev els.Sev eral slicesw eretested,thoughtheauthordidnotindicateexactlyho wman y ,norw asa separationoftrainingandtestingdatamade.Resultsw erequalitativ elybasedon comparingthesystem'ssegmen tationwiththeoriginalra wdata. Sonk a,T adik onda,andCollins[104]o v er-segmen tT1-w eigh tedimagesusing edgebasedregiongro wingwithoperatoradjustedsmoothingparameters.Agenetic algorithmthenin terpretstheedgestructurestopro videnalsegmen tation.Aset of28normalslicesw asa v ailable,witht w en t yoftheseslicesrandomlyselectedfor trainingandtheremainingeigh tslicesastesting.Segmen tationaccuracyw asjudged b yareaandlabelingerrorsofsev en teenneuroanatomicstructures,witherrorsfound toberelativ elysmall.Theauthorsac kno wledgethelimitationsof apriori kno wledge aboutbrainstructures,positions,andrelationshipswhenconsideringpathological

PAGE 67

52 cases.Geneticalgorithmsw ereusedb yV elth uizen[124]toextractthebestfeatures andndaglobalsolutionfortheoptimizationfunctionofaparticularclassier,suc h asF CM.An um berofoptimizationfunctionsw erealsotested,suc has J m andcluster co v ariance. Multiplescansfromv epathologicalsubjectsw ereexaminedandthe geneticalgorithmbasedmethodimpro v edperformanceinF CMsegmen tations,but notwithotherclassiers. Afuzzyrule-baseisusedb yNamasiv a y am[87,89,88]topresegmen tnormal m ultispectralslices.HistogramsareanalyzedinT1,PD,andT2featureimagesand fuzzysetsarecreatedtorepresen tspecictissues.Pixelswithhighmem bershipsina particulartissuearelabeledandusedtoinitializethesemi-supervisedF CM[4]algorithmtoclassifytheremainingunlabeledpixels.Edgedetectionbasedmethodsare usedtow eigh tthePDhistograminabnormalsliceswithreducedtissuecon trast.The fuzzyrulesw eregeneratedfromatrainingsetofsixnormalslicesfrom5v olun teers and4abnormalslicesfrom4patien ts.Thesystemw astestedon33slicesfrom8v olun teersand62slicesfrom7patien ts.UsingkNNsegmen tationsasapseudo-ground truth,thefuzzysystemw asev aluatedb ycomparingitssegmen tationswiththoseof thesamesliceusingstandardF CM.Ov erall,thefuzzysystemproducedcomparable resultstothekNNmethod,butrequiredonly one-fth thecomputationtime.The fuzzysystem'sresultsarealsocomparabletosegmen tationsgeneratedb ythiskno wledgebased-systemin[18].Asimilarrule-basedsystemw asoriginallyin troducedb y Chang,Hillman, etal. [14,15,50].Then um berofslicestestedw aslimited,ho w ev er, consideringonlyasingleh umanMRimageandratbrainimagerespectiv ely Dellepiane,V en turi,andV ernazza[31]usedfuzzyimagemodelingtodescribe tissue-specicfeatures.F uzzymem bershipsetsco v ersliceinformationsuc hastissue gra ylev els,regionshapesandsize,andgeneralanatomicalstructures.Anatomical informationisstoredinahierarc h ysimilartoaseman ticnet.Additionalanalogical andpropositionalkno wledgeisusedtomapsegmen tedregionstolabels.Themodels

PAGE 68

53 forpathologicalslicesw eregeneratedfromatrainingsetofthreeslices,withregions andstructuresofin terestman uallyassignedtotheproperclass.Anadditionalslice w asusedfortesting. Aconfusionmatrixforthetestcasew ascreatedbet w eena man ualsegmen tationandthesystem'ssegmen tation.Resultssho wnoerrorinlesion segmen tation,butthelesionsegmen tationgeneratedb ytheirsystemdidnotremo v e surroundingextra-cranialtissue. Errorw asalsofoundbet w eenotheranatomical structures. Kno wledge-basedimagemodelsw erealsousedb yH.Li,Deklerc k,and etal. [72] tosegmen tCTimages.A\domainblac kboard"iscreatedtoholdinformationprimitiv esgatheredb ylo wlev elimageprocessingtools. Analogicalandpropositional kno wledgemaptheimageprimitiv estobrainmodelobjects.CTv olumesw ereacquiredfromeigh tnormalsubjectswitheac hv olumecomprising14slices.Fiv esets w ereusedastrainingwhiletheremainingthreesetsw eresa v edfortesting.Resultsare v alidatedb ycomparingtheautomaticsegmen tationofmajoranatomicalstructures againstman uallylabeledsegmen tationsofthesamestructures.Dieren tstructures ofin terestw erefoundindieren tslices,socomparisonsw eremadeonaperv olume basis.F orthetrainingset,amaxim umof5 : 94%w asfound,whilethetestsethad upto7 : 82%error.Onlynormalsw ereprocessed,ho w ev er,andtheauthorsdonot discussho wsuc hasystemw ouldhandleabnormalcases. MenhardtandSc hmidt[82]com bineaframe-basedkno wledgebaseandan inferenceenginetoautomatein terpretation(labeling)ofanatomicalregions. The processofsegmen tationisdividedin toaseriesofsub-stepswiththeappropriate kno wledge,suc hanatomicalconstrain ts,location,andhistogramdistributions,includedtoguidetheprocess.Segmen tationsarethen\in terpreted"inseman ticterms thataremeaningfultoradiologists,thoughnodirectv alidationofthesein terpretationsisoered.Thesystemw asdev elopedontextualandpictorialdatafromsubjects diagnosedwithwithMSorbraintumors.Menhardtalsocreates\iconicfuzzysets"

PAGE 69

54 in[81]thataresimilartothefuzzypartitionmatrix U inF CM,butalsocon taining moreabstractdomainkno wledge.The\iconicfuzzysets"arethenconsideredb ya kno wledgebasethatclassieseac hMRpixelindividually ,ratherthanasacluster. Lo w-lev elpropertiesem beddedwithinarule-basedsystemw erealsoimplem en tedb y NazifandLevine[90],thoughforrobotvision,notMRimages,b yusingbasicimage processingoperationstoextractimagepropertiesforanexpertsystemtoanalyzeand labelR OI's. 3.2.3 Statistical/P arametricMethods Muc hoftheearliestw orkinMRsegmen tationhasbeenintheareaofstatistical orparametricmethods. V annierandhisgroupin[120,119]comparedan um ber ofsegmen tationmethods,frommaxim umloglik elihoodtounsupervisedclustering, buttheirprimarygoalw astheunderstandingoftissuedistributionsinaparticular pulsesequence,suc hasthoselistedinT able2. Theynotein[119]acomputer's abilit ytoac hiev eunsupervisedpartitions,in35patien tcases,thatcorrespondw ell withsubjectiv eev aluationsofexpectedtissuedistributions.Gerig,Martin,Kikinis, etal. [38,39]alsofocusedonthedistributionandseparabilit yoftheMRpixelsin featurespace,mostnotablyho wadv ancemen tsinMRscannersha v eledtotigh terand morecompactclusters.Theauthorsadv ocateunsupervisedmethods,whenoptimal imagedataisa v ailable,sincetheydonotrelyonman ualin terv en tion/trainingand oerreproducibilit y Fletc her-Heath,Barsotti,andHornak[35]drewman ualboundariesbet w een tissuet ypesin20v olun teers,rangingfrom17to72y earsinage,toallo wtheir distributionstobeanalyzed.Theirgoalw astondarangeofcalculatedT1,PD, andT2v aluesthatbestdenedtissueclusterboundaries.Segmen tationw asac hiev ed b yin teractiv elymanipulatingthesecalculatedv aluesfromaspecictissuecluster

PAGE 70

55 (locatedb yexaminingthehistogramsofin v erseT1andT2featureimages)un til optim umclassicationw asfound.Althoughnogroundtruthw asusedforcomparison, theauthorsfoundahighcorrelationbet w eenv olun teersforthelocationsofmajor tissueclustersinfeaturespace,locationswhic hw ereconsisten twiththekno wledge describedinSection2.2. Statisticalestimationsoftissuedistributionsha v ealsobeenappliedto w ards segmen tation.T axt,Lunderv old, etal. segmen teduterinecorpustumorsin[112]. Ac-meansclusteringalgorithminitiallypartitionedanMRimageofafemaleuterus in to20to40clusters,withclustersbelongingtothesametissueman uallygrouped together. Probabilit ydensit ymodelsofmajortissuesw erethenconstructedfrom thesegroupings.Classicationbasedonthesedensitiesw asdoneusingaBa y esian framew ork. Theprobabilit ymodelsw ereconstructedfromasetofninehealth y v olun teers,fourmaleandv efemale,andsev enfemalepatien tcases.Threedieren t uterinetumort ypesw erefoundinthetrainingset,andaprobabilit ymodelw as constructedforeac h.Thisallo w edthesystemstopredictthespecict ypeoftumor foundinnineadditionaltestcases,asw ellactuallysegmen tthetumorfromthe image.Thetumort ypepredictedb ym ultispectralanalysisw asgenerallyconrmed b ysubsequen thistopathologicalexamination. Largerdegreesoferrorw erefound, ho w ev er,whencomparingtheactualsizes(largestdiameterinmillim ete rs)ofthe tumors.T axtandLunderv oldalsotak easimilarapproac hinexaminingMRimages ofthebrainin[110,111]usingv eMRc hannels.Correctclassicationratesoftissues w assho wntobebet w een71%and90%.Onlyasingletrainingandtestimagew ere usedrespectiv ely ,ho w ev er,sotherobustnessoftheprobabilit ymodelsisunkno wn. Kam ber,Shingal, etal. [57]usea3Dgeometricprobabilit ymodelofbrain tissuestosegmen tMSlesions.Probabilitiesarebasedonacom binationofpixelintensit yandlocationwithinthebrainandw ereoriginallygeneratedfromatrainingset of12v olun teers.Tw elv epatien tswithMSlesionsw erethenexaminedusingav ariet y

PAGE 71

56 ofclassiers,includingminim umdistance,Ba y esian,andprunedandunpruneddecisiontrees.Theirrespectiv eMSlesionsegmen tationsw erecomparedagainstman ually labeledgroundtruth(within tra-observ erv ariabilit yof3-5%).ArticialMSlesion phan tomsw erealsousedtomak ecomparisonswhilea v oidingobserv erv ariabilit y .In termsoferrorrate,theBa y esianclassierperformedthebest,thoughallfourmethods hadlessthan5%error.Acom binationofcon tourtracingandstatisticalanalysisw as attemptedb yHolden,Steen,andLunderv old[51].T obuildtheclassier,parametric estimationsofthemajortissueclassesw eregeneratedfromasetofcalibratedimages from11dieren ttrainingsubjectsand15slicesofasinglemetastatictumorMR v olume.Dicultiesw eresho wninproperlysegmen tingthetumorandtheauthors suggestremediesforthisproblem. W ells,Grimson,KikinisandJolesz[128]useav arian tofthemaxim umlik elihoodmethodcalledexpectationmaximiz ation(EM)theycall\adaptiv esegmen tation."Kno wledgeoftissuepropertiesandwithinslicein tensit yinhomogeneitiesare usedb ytheEMalgorithmtocorrectMRimageinhomogeneitiesandenhanceseparationoftissuesinfeaturespace.Theactualsegmen tationprocessisthenperformed b yanautomaticnonparametricclassier.Consideringsliceswithin traandin ter-scan inhomogeneities,signican timpro v em e n tinsegmen tationqualit yissho wn.Theauthorsstatethemethodhasbeenused\eectiv ely"ino v er1000brainscans,butdonot indicateho wman yslicesw ereusedtogeneratetheprobabilit ymodelsorho wman y slicesw erenormal/pathological,makingcomparisonswiththekno wledge-basedsystemheredicult.OnlyasinglesagittalPDslicew asquan titativ elycomparedwith man uallylabeledR OI's(ofwhiteandgra ymatteronly)fromv eraters.TheseR OI's w ereobtainedb yselectingpixelswhoselabel(whiteorgra ymatter)w asthesame inatleastfourofthev eman ualsegmen tations.A19-23%dierenceisreported, consisten twithin ter-observ erdierences.Theresultsofasupervisedsegmen tation

PAGE 72

57 tec hnique(b ythesameobserv ers)isalsoincluded.Theadaptiv emethodw asfound toagreemorecloselywiththeman uallylabeledimages. AnEMalgorithmisalsousedb yLunderv oldandStorvik[80]tosegmen tbrain parenc h ymaandCSFv en tricles.T rainingv o xelsareacquiredviaac-meansclustering ofasingletrainingslice,andthetissueparametersarethenestimatedb ytheEM algorithm.Modelsofdynamiccon toursstoredinaBa y esianframew orkareusedto performtheactualclassication.T enadditionalslicesfromtenpatien tsw ereused fortestingpurposes,andsatisfactorysegmen tation(basedonvisualinspection)w as ac hiev edinamajorit yoftheslices.F ailuresoccurredwhenmodelassumptionsw ere violated,eitherb ythepresenceofpathology ,signalabnormalities,ortheshapesof thev en triclesdidnotmatc h.Also,modelassumptionslimittherangesofslicesthe methodcanbeappliedto. Cohen,Andreasen, etal. usedsupervisedR OI'stocreatelineardiscriminan t functionsinPDandT2-w eigh tednormalslices[27].Theauthors,ho w ev er,focusmore ontheeectsofobserv erv ariabilit y ,reproducibilit ybet w eenthesameanddieren t MRcoils,m ultipletrainingclasses,andotherfactors. F orexample,in ter-observ er reliabilit yw astestedb yha vingt w oindependen tobserv ersselecttrainingregionsfor thesame10subjectimagesetandfoundtoha v ebet w een79-96%agreemen tformajor braintissues,whilein tra-observ eragreemen tw asratedat61-96%. Johnston,A tkins,Mac kiewic h,andAnderson[56]useastoc hasticrelaxation basedmethodcalled\iteratedconditionalmodes"(ICM)tosearc hforMSlesions. TheICMmethodissupervised,usingman uallydra wntissueR OI's(withineac hslice) toguideit,andcanoperateoneitheraPDorT2-w eigh tedimage,orboth.Aslice's ra wdataisrstpre-processedtocorrectMR-coilinhomogeneitiesandextractthe in tra-cranialregionfromtherestoftheimage.TheICMmethodistheniterativ ely appliedtorstsegmen twhitematter(where95%ofallMSlesionsoccur)fromthe restofthebrain,thentheMSlesionsfromwhitematter.AftertheICMalgorithm

PAGE 73

58 isapplied,post-processing,intheformofmorphologicaloperations,isappliedto renetheMSlesionboundaries. Dual-ec ho(PDandT2-w eigh ted)slicesfromv e MRv olumes,with21-27slicesfoundineac hv olume,w ereprocessedandasimilarit y indexisusedforsystemev aluationb ycomparingthesystem'slesionsegmen tation toaman uallesionsegmen tationforthesameslice.V aluesrangedbet w een0 : 2(v ery poor)and0 : 8(excellen t)withana v erageof0 : 505.Theauthorsnotethatslicesthat w eresegmen tedwithbothaPDandT2-w eigh tedimageperformedm uc hbetterthan sliceswhereonlyaPD-w eigh tedimagew asa v ailable. Karssemeijer[58]alsousesthismodel,thoughhissystemsearc hesforabdominaltissuesinCTandx-ra yscans.Errorw asratedatlessthan7%iniden tifying12 majorstructuresofin terestwhencomparedwithman uallysegmen tedimages.The authordidnot,ho w ev er,indicateho wman yimagesw eretested.Alineartransformationusingoperatorselectedtrainingexemplarsw asproposedb yKohn,T anna, and etal. [65].Someha v enotedthatsuc hamethodw orksw ellfornormals,where tissuedistributionsarestable,butperformspoorlywhendealingwithpathological cases[122]. Infact,thegreatestobstaclewithparametricmethodsisgenerating reliableprobabilit ydensit ymodelsforpathologicalslices. 3.2.4 NeuralNet w orks Someresearc hersha v eturnedtotheuseofarticialneuralnet w orksasameans oflearningthec haracteristicsofMRdatathatma ybetoocomplextocapturewith clusteringorstatisticalmethods.Li,Bhide,andKabuk auseabooleanneuralnet w ork (BNN)toimplem en tav ariationofc-meansclusteringtopro videaninitialsegmen tation[73].Imagebasedkno wledgeisthenin tegratedin toaconstrain tsatisfyingBNN tomergeandlabeltheregions.Theirh ybridsystemissho wncomparefa v orablyto

PAGE 74

59 aHopeldneuralnet w orkandtraditionalsim ulatedannealingandac hiev essegmentationinfew eriterations.Onlyt w onormalMRimagesw eretested,ho w ev er,and themethodrequiredoperatorselectedtrainingv ectors,sothemethod'srobustnessis unkno wn. Ozk an,Da w an t,andMaciunas[93]includedx-ra ycomputertomograph yas anadditionalfeatureimageinafeed-forw ardnet w orktrainedwithpixelscorrectly labeledb yapretrainedclassier.Abac k-propagatingnet w ork(BPN)iscompared withamaxim umlik elihoodclassier(MLC)onMRandCTslicesfromnormals (then um berw asnotspecied)andt w ometastatictumorpatien tswitheigh tand fourslicesrespectiv ely .T rainingR OI'sw ereman uallyselected.F ornormalslices, theBPNconsisten tlysho w edless\noise"thantheMLCandresultssho w edthe BPNw asabletomoreeectiv elygeneralizeonasmallertrainingset.P athological casesw erecomparedwithman uallylabeledgroundtruthimagespro videdb ythree dieren t\experts"andev aluatedusingasimilarit yindex(denedinSection6.3.4). TheBPNandMLChadana v eragesimilarit yindexof53and46respectiv ely ,while in ter-ph ysiciansimilarit yw asmeasuredat55.Theauthorsac kno wledgethedicult y insegmen tingMRv olumesasasingledataset,statingthatthemeanin tensit yand standarddeviationc hangessignican tlywhilemo vingthroughthev olume. Kisc hell,Keh tarna v az, etal. comparedfourneuralnet w orksinconstructinga neuralnet w orkbasedexpertsystem[61].Thesupervisedbac k-propagation(BPN) andcoun ter-propagation(CPN)net w orksw ereexaminedalongwiththeunsupervised Kohonennet w orkandanalogadaptiv eresonancetheoryAR T2.Thedatasetused consistedoffourcoronalslicesfromeac hof16c hildren,4normal,and12withmild tosev ereheadinjuriesthatresultedineithermacrocysticencephalomalaciaorgliosis lesions.Dual-ec hodataw asused,butfromthePDandT2images,atotalofnine featuresw ereextracted,includingsignalunion(PD+T2)anddierenceimages(PDT2),asw ellasenergyanden trop ybasedtextureimages. Pixelsfromeac htissue

PAGE 75

60 ofin terestw ereman uallyselectedfromthedatasettotrainthenet w orks,withthe remainingpixelsintheimagesactingastestdata.Ofthefournet w orkstested,the BPNmethodga v ethebestresults,basedonvisualinspection.Whencomparedto expertobserv ers(usinganin teractiv esegmen tationmethod)forthediscrimination ofbraintissuesinnormalcases,a6%estimateofdiscrepancyw asfoundbet w een thenet w orkandh umanobserv ers.TheBPNnet w orkalsosuccessfullydetectedthe t w ot ypesofpathologicallesionsfromnormalbraintissues,butdirectcomparisonsof lesionmeasuremen tw erenotmade.Alirezaie,Jernigan,andNahmias[1]comparea learningv ectorquan tization(L V Q)net w orkagainstbac k-propagationinsegmen ting normalandabnormalMRv olumes.T rainingpixelsareselectedinbothmethodsand theauthorsstatethattheL V Qmethodisfasterandmorerobustacrossan um berof v olumes,thoughtheydonotspecifyho wman yv olumesthatis. Someresearc hersha v etak enanunsupervisedapproac htousingneuralnetw orks. AHopeldneuralnet w orkw asusedb yZh uandY an[132]todetermine tumorboundariesinT1-w eigh tedaxialMRimagesofthebrain.Afterpre-processing thev olumesliceswithalo w-passlter,a\rstslice"isman uallyc hosen,basedon thetumor'scon trast(withinaparticularslice)tosurroundingbraintissues.Arough segmen tationisgeneratedb yapplyingathresholdsettoaT1signalin tensit ygreater thanthatofthewhitematterintheimage,thoughtheauthorsdonotspecifywhether thisisautomaticallyorman uallydone.Imagemorphology(erosionanddilation)remo v esallregionsexcepttumorandcreatesaninitialestimateofthetumor'scon tour. Thisinitialcon touristhenpassedtotheHopeldnet w orkforrenemen t.Kno wledge ispropagatedbet w eenslicesb yusingthenaltumorboundaryofapreviousslice toinitializesegmen tationofacurren tslice. Tw oMRv olumesareprocessed,but onlysliceskno wntocon taintumorareconsideredandtheautomaticsegmen tations arenotcomparedwithan ygroundtruth.F urthermore,onlyT1-w eigh tedimagesare sho wn,soitisunclearifm ultispectraldataisused.

PAGE 76

61 AHopeldnet w orkw asalsousedb yAmartur,Piraino,andT ak efuji[2]to segmen tdual-ec ho(PDandT2-w eigh ted)images.Theirnet w orkw asbasedonan energyminimi zationfunctiondesignedtoallo wforh yper-ellipsoidalclusterdistributionsandslices(asingleimagefromanormalv olun teerandpatien tcaseeac h) w ereprocessedm ultipletimes,eac hforadieren tn um berofclusters.Segmen tation qualit yisbasedonvisualinspection.Lin,Cheng,andMaocom bineneuralnet w orks andfuzzyclusteringin[76].Acompetitiv elearningnet w orkthatemplo ysa\winner tak eall"sc hemeisin tegratedwithafuzzyclusteringalgorithm.In[75],theycreated anunsupervised2DHopeldnet w orkthatusestheF CMalgorithmtoeliminatethe needforndingw eigh tingfactorsintheenergyfunction.

PAGE 77

62 CHAPTER4 DETECTINGP A THOLOGYINSLICESBELO WTHEVENTRICLES AsstatedinChapter3,attemptstoclusteranMRv olumeasawholeproduced poorresultsduetodatainhomogeneit y[121]. Sincewithinsliceinhomogeneit yis acceptablewiththeG.E.scannerandprotocolusedinthisw ork,asnotedinAppendix A,slicescanbeseparatelysegmen tedandprocessed.F ollo wingthestageslistedin Figure9,anMRsliceisinitiallysegmen tedb yF CMandthesofttissuesofthebrain areextracted. Thein tra-cranialregionisthencomparedwithpreviouslydened qualitativ emodels.Ifasignican tdeformationisdetectedwithinatissue,theslice isclassiedasabnormal. Otherwise,thesystemclassiesthesliceasnormaland proceedstothenextslice(mo vingdo wnthehead)un tilallslicesinthev olumeare processed. Thisiscalledautomaticv olumeprocessingbecausethereisnoexternalinteractiononcethev olumeprocessinghasstarted.Infact,thesystemhasonlythe follo wingrequiremen ts:(1)theinitialslicehasalreadybeenprocessedandtherst sliceconsideredb ythissystemliesimmedi atelybelo wit;(2)pathologyw asenhanced withaparamagneticcompound;(3)v olumeslicesareprocessedinsequen tialorder toallo wkno wledgetobepropagatedfromoneslicetothenext;(4)theheadw as positionedinthecoil(subjectlookingstraigh tup)inauniformmanner,thoughnot inanexactposition.

PAGE 78

63 4.1 StageZero:InitialSegmen tation Giv enanMRslicedataset,therststepistogenerateaninitialsegmen tation. Allslicesaredeliberately\o v er-clustered"b ytheF CMalgorithmin tomore classesthantissuet ypesafterexperimen tsb yLiin[70,71]sho w edthatusinga one-to-onecorrespondencebet w eenthen um berofclassesandtissuet ypesresulted inpixelsbelongingtodieren ttissueclassesoftenbeingplacedin tothesamecluster,especiallyinsliceswithsignican tpathology .Ov er-clusteringresultsinalev el ofo v er-segmen tation,wherethesametissuet ypeissplitin tot w oclusters,butit reducestheprobabilit yanddegreeofmixedclasses[70,71]. Ov er-clusteringdoes notguaran teeitsprev en tionofunder-segmen tation,ho w ev er,especiallyforobjects thatarediculttoseparate.Ov er-clusteringma yinsteado v er-segmen tobjectsthat w ereproperlyclusteredin toasingleclasswhenusingfew erclusters.Moreo v er,an increaseinthen um berofclassesusedforclusteringwillincreasetheamoun tofcomputationtimelinearlyinthen um berofclusters.Li[70,71]empiricallyfoundthat usingtenclustersresultedinsatisfactorysegmen tationwhilek eepingcomputational costsdo wn. Onceasliceissegmen tedwithF CM,kno wledgediscussedinSection2.2can beappliedenablingthedistributionoftheclustercen terstobematc hedwithcorrespondingtissuelabels.Figure12sho wsat ypicalexample.Sixc haractersareused torepresen tclassesofairandbone(A),extra-cranialtissues(B),whitematter(W), gra ymatter(G),pathology/tumor(P),andCSF(C).Figures12(b)and(c)sho wbetterillustrationsb yprojectingtheclustercen tersin toT1andT2spaceandin toonly theT2spectrumrespectiv ely .Byexploitinga\hierarc h y"infeaturespace,through acluster'sparticular\rank"(inascendingorder)inaparticularfeature,tissuesof in terestcanbemoreeasilylocated.Onceapossibletissuet ypehasbeenlocated,it ma ybev eriedandanalyzedbasedonexpectedanatomicalproperties.Itshouldbe

PAGE 79

64 High T2 Low T1 High PDC G G W B B B B A PLow PD High T1 Low T2(a) A B B B B W G G C High T1 High T2 P Low T1 Low T2(b)Air Fat, Skin, Muscle White Matter Gray Matter High T2 Signal Intensity Low T2 Signal Intensity Pathology CSF (c) Figure12.ClassCen tersforanAbnormalSlice:(a)T1,PD,andT2Space,(b)in theT1,T2Plane,and(c)intheT2Spectrum.A=AirandBone;B= Extra-cranialtissues ;W=whitematter;G=gra ymatter;C=CSF;P =P athology notedthatthishierarc h yisbasedontheclusters'relativ eorderinfeaturespaceand notan yparticularv alue.Thisgiv estheextractedkno wledgegreaterexibilit ysince itisnotbasedonan yparticularF CMpartition. 4.2 StageOne:CreatingtheIn tra-CranialMask Inordertoaccuratelydetectabnormalitieswithinthebrain,analysism ustbe limitedtopixelsthatcorrespondtothein tra-cranialregion,whic hcon tainthebrain's softtissues.Pixelsbelongingtoextra-cranialtissues,suc hasair,skin,andfatare notofin terest.Therefore,giv enaninitialF CMsegmen tation,thegoalofStageOne istocreateanin tra-cranialmaskthatcon tainsonlypixelsbelongingtosoftbrain tissues.Onceanin tra-cranialmaskisgenerated,asliceT emplatema ybematc hed toitandpathologydetectionma ybeperformed.

PAGE 80

65 (a) (X ,0) r Origin Gap (b) (c) Figure13.SeparatingExtraandIn tra-CranialClusters.(a)Initialsegmen tedimage withairalreadysettobac kground(white);(b)aquadrangleo v erlaidona binaryimageof(a)aftersettingthelo w estthreeT2classestobac kground (white)andtheremainingclassestoforeground(gra y);(c)classesthat passedquadrangletest. 4.2.1 ExtraandIn tra-CranialClusterSeparation Assho wninFigure12,afteraslicehasbeenpartitionedb ytheF CMalgorithm, eac hclusterwillcon tainpixelsbelongingprimarilytothesametissuet ype.Clusters con tainingprimarilyextra-cranialpixelsarenotofin terestandshouldberemo v ed fromfurtherconsideration.Clusteringwilloccasionallyplaceanextra-cranialtissue classbet w eenwhitematterandgra ymatterinT2space,assho wninFigure12(c), prev en tingasimpleT2thresholdfrombeingused.Li[70,71]usedanatomicalkno wledgethatextra-cranialtissuessurroundthebrainandarenotfoundwithinthebrain itself.Therefore,ifagrossestimateofthebrainiscreated,thenclustersconsisting ofextra-cranialtissueswillha v ev eryfewpixelsinsidethisestimatedbrain,while clustersofin tra-cranialtissueswillha v easignican tn um ber. Creationoftheestimateofthein tra-cranialregionbeginsb ynotinginFigures12(b)and(c)thatthelo w estthreeclassesinT2space,aftersortinginascending order,belongeithertoairorakno wnextra-cranialtissue.Thesethreeclassescan automaticallybesetto\bac kground"(whic hcon tainpixelsnotofin terest)anda

PAGE 81

66 binaryimagecanbecreatedfromtheinitialsegmen tationsettingthetheremaining sev enclassestoforeground.Figure13(a)sho wsat ypicalsegmen tationfromF CM, withclustersbelongingtoairalreadysettobac kground(white).Abinaryimage, Figure13(b),canbecreatedb ysettingtheremainingclusterstoforeground(gra y), andsho wsavisiblegapbet w eenextra-cranialtissuesandthetissuesofin terest. Thecen troidofthebinaryimageiscalculatedtoappro ximatethecen terofthe head[52],whic hisusedastheoriginofaCartesiancoordinatesystemthatiso v erlaid onthebinaryimage,assho wninFigure13(b).F romtheorigin,thesystemmo v es righ talongtheX-axisfromun tilapoin t( X r ; 0)isfoundsuc hthatan8 8windo w from( X r ; 0)to( X r +7 ; 0)andcen teredaroundtheXaxishas K 25bac kground pixelsinthewindo w,with L 5bac kgroundpixelsinasinglero w.Thepoin t( X r ; 0) isreferredtoastherigh tinnerpoin t.Aleftinnerpoin t( X l ; 0)canbefoundsimilarly ascantop(0 ;Y t )andbottom(0 ;Y b )innerpoin ts,exceptthattheyoperatealongthe YaxisandtheLbac kgroundpixelsm ustbefoundinasinglecolumninsteadofa ro w. Thesefourinnerpoin tsproduceaquadrangleassho wninFigure13(b)and pro videtheappro ximatedin tra-cranialregion.T omeasurewhic hclustersha v easignican tn um berofitspixelswithinthequadrangle,eac h Cluster i (exceptthelo w est threeT2clusters,whic hw erealreadymark edasextra-cranialincreatingthebinary image)hasa\quadranglepercen tage"( QP i )calculatedb y: QP i = Num berofpixelsof Cluster i withinquadrangle. T otaln um berofpixelsin Cluster i (4.1) TheQPofaclusteriscomparedagainstathresholdsetaccordingtothecluster's rankinT2spaceandtheT emplateofthepreviousslice,sho wninT able4.2.1.The thresholdv aluesusedw eresetb yexaminingthea v ailabletrainingslicesanddeterminingthemaxim umQPv aluesofkno wnextra-cranialtissuesandtheminim umQP

PAGE 82

67 T able3.RulesforClusterSeparation. Ifthepercen tageofthecluster'spixels withinthequadrangleislessthanthethreshold,theclusterismark edas extra-cranial.Otherwise,itismark edasin tra-cranial.ClusterswithaT2 rankingof3orlessareautomaticallymark edextra-cranial. Cluster'sRankinT2Space T emplateofPreviousSlice %Threshold Rank 8 1,4,or5 6% Rank 7 1or4 13% Rank 7 5 10% v aluesofkno wnin tra-cranialtissuesand\splittingthedierence."Inslicesthatdo notin tersectthev en tricles,clusterswithahighT2rank(8+,CSFandimm ediatel y neigh boringgra ymatter)arefoundalongtheperipheryofthebrainandareexpected toha v ealo w erQPv alue.Anadjustmen tw asalsoneededforT emplate5slicessince asignican tamoun tofbraintissueisfoundinthetemporallobes,outsidethequadrangle.Ifthecluster'spercen tageislessthanthethreshold,theclusterismark edas extra-cranial.Otherwise,itismark edin tra-cranial.Therstmomen tofthehead andtheQPv aluesofeac hclusteraresa v edforuseinlaterprocessingsteps. 4.2.2 Reco v eringIn tra-CranialClustersinT emplate5LSlices AfterapplyingthequadranglerulesinT able4.2.1,theclustersfromtheF CM segmen tationha v ebeendividedin toasetofextra-cranialtissueclasses,collectiv el y referredtoasGroup1clusters,andasetofin tra-cranialtissueclasses,kno wnas Group2clusters. Thereshouldbeaminim umofthreeGroup2clusters,onefor whitematter,gra ymatter,andCSFrespectiv ely .IflessthanthreeGroup2clusters arefound,thenadeformationofthequadranglehasoccurred,createdwhenone orbothofthetemporallobesbegintoseparatefromthecerebellum,assho wnin Figure14(b). ThisindicatestheslicebelongstoT emplate5Landifnopathology w asdetectedinthepreviousslice,processingwillhalt.

PAGE 83

68 (a) (b) (c) (d) Figure14.Reco v eringIn tra-CranialClusters.Giv enaninitialsegmen tation(a)and adeformed\quadrangle"(b),in tra-cranialclusterscanbereco v eredb y o v erla yingthein tra-cranialmaskfromtheT emplate5slicespatiallyclosesttothenec k(c)anddetectingwhic hclustersha v easignican tn um ber ofpixelswithinthemask,assho wnin(d). Ifpathologyw asdetectedinthepreviousslice,thein tra-cranialclustersare reco v eredusinganewappro ximationofthein tra-cranialregion.Kno wledgeoftemplateorderingdescribedinSection2.3indicatesthatT emplate5slicesimm ediatel y precedeT emplate5Lslicesandha v ew elldenedin tra-cranialmasks. Byaccessingthein tra-cranialmaskfromtheT emplate5slicespatiallynearesttothenec k (the\lo w est"slicepositiv elyiden tiedasT emplate5),anewappro ximationofthe in tra-cranialregionisa v ailable. Oncethemaskofthelo w estT emplate5sliceisiden tied,foreac h Cluster i withaT2rankof4orgreater,aratioiscalculatedb ydividingthen um berofpixels in Cluster i con tainedb ythemaskb ythetotaln um berofpixelsin Cluster i All clusterswitharatioof30%orgreateraremark edasGroup2,withtherestmark ed asGroup1.AnexampleofthisprocessisillustratedinFigure14. 4.2.3 DetectingF alseWhiteMatter Withallextra-cranialclustersremo v ed,thedistributionsinFigure12indicate thatwhitemattershouldbetheGroup2clusterwiththelo w estT2v aluecen troid. Insomeslices,ho w ev er,protectiv emeningialtissuesco v eringthebraincanbefound

PAGE 84

69 withinthessuresofthemajorlobes.Asaresult,aclustercorrespondingtothese tissuescanpassthequadrangletest,falselyclassifyingitasGroup2,anddisplace whitematterasthelo w estT2v alueGroup2cluster. These\falsewhitematter" clusterscanbedetectedb yapplyingkno wledgeofexpectedrelativ eMRdistributions infeaturespace,assho wninFigure12. AssumeaGroup2cluster, Cluster i ,where i indicatesacluster'srankinascendingT2space,andfeaturev alues h T 1 i ;PD i ;T 2 i i .Giv en Cluster i +1 andfeature v alues h T 1 i +1 ;PD i +1 ;T 2 i +1 i ,thefollo wingtrendscanbeusedasheuristics: 1. T 2 i +1 >T 2 i ,giv ensincetheclustersareorderedinT2space. 2. PD i +1 >PD i ,thoughnotalw a yswithsomewhiteandgra ymatterclusters. 3. T 1 i +1 PD 1 AND T 1 2 >T 1 1 b ymo rethan T 1 1 5% OR T 2 2 >T 2 1 b ymo rethan T 2 1 60% AND QP 1 < 0 : 30 THEN Ma rk Cluster 1 as\falsewhitematter" Conrm Cluster 2 aswhitematter ELSE Conrm Cluster 1 aswhitematter TheruleassumestherstheuristicsincetheclustersarealreadyorderedinT2space. ViolationsofthesecondheuristicconcerningPDoccursprimarilyinsliceswhere whitematterhasbeen\split"in tot w oadjacen tclusters.F alsewhitematterclusters, ho w ev er,ha v estrictlylo w erPDv aluesthantruewhitematterclusters.Therefore, requiringthat PD 2 >PD 1 prev en tstherulefromexecutingonsliceswithsplitwhite matter.Finally ,becausefalsewhitematterclustersareextra-cranial,theygenerally

PAGE 85

70 (a) (b) (c) Figure15.CreatingtheIn tra-CranialMask.Giv enaslice'sGroup2clusters(a),b y creatingabinaryimage(b)oftheseclusters,imagemorphologyandan eigh t-wiseconnectedcomponen tsoperationcanextractthein tra-cranial region(c). ha v erelativ elylo wQPv aluesincomparisontotruewhitematterclusters.Therefore, c hec kingtheQPv alueactsasanadditionalsafeguardtok eeptherulefrompossibly executingonatruewhitemattercluster.Ifaclusterisfoundtobefalsewhitematter, itisremo v edfromGroup2andplacedin toGroup1. 4.2.4 ExtractingtheIn tra-CranialRegion OnceGroup2clustersha v ebeeniden tied,aninitialestimateofthebrain canbegeneratedb ysettingallGroup1clusterstobac kground.Figure15(a)sho ws thebrainalongwithsome\noise"belongingtoextra-cranialpixelsmisplacedin toa Group2clusterduringF CMsegmen tation.Mostofthesenoisypixelsarerelativ ely sparsespatially ,butsomeslicescancon tainlargerregions,suc hastheey es,assho wn inFigure15(a).Anatomically ,ho w ev er,thebrainisconsisten tlythelargeststructure intheimageandcanbeextractedb yrstsettingallGroup2clustersinFigure15(a) toforegroundtocreateabinarymask(Figure15(b)).Thebinaryimageiseroded witha5 5operatortoremo v esmaller\noise"regionsandenhanceseparationof thebrainfromsurroundingtissues.Aneigh t-wiseconnectedcomponen tsoperation

PAGE 86

71 follo wsandthespatialcomponen twiththelargestn um berofpixelsispreserv ed,with allotherregionssettobac kground. A5 5dilationreco v erspixelslostduringtheerosionoperation.Holes,pixels correspondingtoin tra-cranialtissuesthatw eremisplacedin toaGroup1cluster duringF CMsegmen tation,arelledusingthemethoddescribedinSection2.7.3.In someslices,smallholesneartheperipheryofthein tra-cranialregionw ere\opened" b ytheerosionoperation.T olltheseholes,another5 5dilationisperformedtoclose thegapscreatedb ytheoriginalerosionoperation,andanotherhole-llingprocess performed. Finally ,a5 5erosionshrinksthemasktoitsoriginalsize,creating anin tra-cranialmaskandcompletingStageOne. Anexamplemaskissho wnin Figure15(c). 4.3 StageTw o:SliceT emplateIden tication Onceanin tra-cranialmaskhasbeencreated,thegoalofStageTw oistoidentifythecorrecttemplatefortheslice.Aiterativ eparadigmisemplo y edinwhic ha templateisrstten tativ elyassignedbasedontheshapeofthein tra-cranialmask.The ten tativ etemplatelabelisthenusedtoguideadditionalmaskrenemen tsaccording toqualitativ emodelstructuresexpectedforthattemplate.Analtemplatedeterminationwillthenbemadeforcertainslicesbasedonshapeanalysisofanatomical structures. 4.3.1 InitialT emplateDetermination Earlytemplatelabeldecisionsarebasedontheassumptionthatthecurren t slicehasthesametemplateasthepreviousslice(locatedimmedi atelyabo v ethecurren tslice)unlessandun tila\transition"bet w eentemplatesisdetected.Atemplate

PAGE 87

72 Top-most Point (X ,Y )tb tbRight-most Point (X ,Y )rb rb Foreground Image Centroid (X ,Y )cb cbBottom-most Point (X ,Y )bb bbLeft-most Point (X ,Y )lb lb(a) (X ,0) (0,Y ) (0,Y ) (X ,0)t lr b(b) (c) Figure16.BoundingBo xes.A\boundingbo x"tsthesmallestrectangle(aligned withthemajoraxes)aroundaforegroundobject,assho wnin(a).Using aboundingbo xonthein tra-cranialmask(b)anditsleftandrigh themispheres(c)returnsinformationtoallo wT emplateiden tication.Image (b)alsosho wsthenew\innerpoin ts"generatedonthein tra-cranial-mask. transitionoccurswhentheshapepropertiesofthein tra-cranialmaskofthecurren t slicedeviatesfromthequalitativ emodelofthetemplateofthepreviousslice.When makinganinitialtemplateassignmen t,somereliablekno wledgeisreadilya v ailable: 1.Asslicesaretak enfromlo w erinthehead,theorderoftemplatetransitionis strictly1 4 5,assho wninFigures5(a),(d),and(e)respectiv ely 2.Thetemplateandqualitativ epropertiesofthepreviousslice, Template previous 3.Therstmomen tofthehead,asdescribedinSection4.2.1. 4.Theshapeofthein tra-cranialmaskofthecurren tslice. Theorderoftransitioncomesfromgeneralkno wledgeconcerningtheanatom yofthe brain,asdiscussedinSection2.3.Thetemplateoftheprevioussliceispropagated kno wledge,men tionedinSection2.6. Ifthecurren tsliceistherst\lo w er"slice beingprocessed,thesliceimme diatelyabo v eitisthe\initialslice"andkno wntobe T emplate1.Therstmomen toftheheadw asgeneratedinStageOne,Section4.2.1, andisreferredtoas( X h ;Y h ). Asstatedearlier,atransitionbet w eent w otemplateshasoccurredwhenthe shapeofthein tra-cranialmaskofthecurren tslicedeviatesfrom Template previous

PAGE 88

73 asdescribedbelo w.Therulesinthissystemarethereforedesignedtodetectthese deviations. Aslice'sshapepropertiesareiden tiedthroughaseriesof\bounding bo xes"[54]andanewsetof\innerpoin ts"(originallypresen tedSection4.2.1).A boundingbo xtsthesmallestrectangle,withitsedgesalignedwiththemajoraxesof theCartesiancoordinatesystem,aroundaforegroundobjectofin terest.Figure16(a) sho wsanexample.Theboundingbo xreturns( X cb ;Y cb ),thecen troidoftheobject. Italsoreturnsthe\top-most"poin t( X tb ;Y tb ),whic histhepoin twiththehighest Y v alue. Similarly ,poin tsforthebottom-most( X bb ;Y bb ),left-most( X lb ;Y lb ),and righ t-most( X rb ;Y rb )poin tsarealsofound,assho wninFigure16(a).Thenewsetof innerpoin tsarecreatedforthein tra-cranialmaskusingthecen troid,( X cb ;Y cb ),asa neworiginoftheCartesianaxes,assho wninFigure16(b). Aftertheboundingbo xisttedaroundthein tra-cranialmask,thehemispheres ofthebrainma yberoughlyseparatedusingthecen troidofthein tra-cranialmask ( X cb ;Y cb ). Thelefthemispherewillbecomprisedofallpixels( X p ;Y p )suc hthat X p
PAGE 89

74 (0,Y )t (X ,Y ) (Left)tb tb(X ,Y ) (Right)tb tb(a)T emplate5Slice (0,Y )t (X ,Y ) (Left)tb tb(X ,Y ) (Right)tb tb(b)T emplate5LSlice OneLobeSeparated (0,Y )t (X ,Y ) (Left)tb tb(X ,Y ) (Right)tb tb(c)T emplate5LSlice BothLobesSeparated Figure17.DetectingT emplate5LSlices. InaT emplate5slice(a),thetop-most poin tsoftheleftandrigh themispheres( X tb ;Y tb ),asdeterminedb yboundingbo xes,mirroreac hother,ha vingappro ximatelythesameYv alueas w ellasanequalandsignican tdistancefromtheY-axis.Thesetop-most poin tsalsoha v easignican tdistancefromthetopinnerpoin tofthe in tra-cranialmask, Y t TheseheuristicsareviolatedinaT emplate5L slice,whenone(b)orboth(c)ofthetemporallobesseparatesfromthe cerebellum. temporallobeshasseparated,ho w ev er,thesequalitativ epropertieswillbeviolated, ascanbeseeninFigures17(b)and(c).Analheuristicusedisthatthecen troid ofaT emplate5Lbrainmaskwillbelocatedev enfurtherto w ardsthebac kofthe brain,a w a yfromthecen troidofthehead.Theseheuristicsareimplem en t edinthe follo wingruletodetectT emplate5Lslices.

PAGE 90

75 GIVENX,YV aluesofT op-MostP ointsof Left/RightHemispheres: ( X Left tb ;Y Left tb ) ; ( X Righ t tb ;Y Righ t tb ) T opInnerP ointofIntra-CranialRegion: (0 ;Y t ) XV alueofT op-MostP ointofIntra-CranialRegion: X ICR tb X,YV alueofCentroidofIntra-CranialRegion: ( X ICR cb ;Y ICR cb ) YV alueofCentroidofHead: Y h IF PreviousSliceisT emplate5 THEN IF min ( Y Left tb ;Y Righ t tb ) 30 OR j X ICR tb )Tj/T17 1 Tf50 0 TD(X ICR cb j < 10 AND j Y ICR cb )Tj/T17 1 Tf49.9999 0 TD(Y h j > 35 THEN Ma rkT emplate5L ELSE Ma rkT emplate5 T ransitionsfromT emplate4to5aredistinguishedb ythecompletedisappearanceofthefron tallobe,transformingthein tra-cranialmaskfroman\o v oid"shape in toa\horseshoe"shape.Thiscausesthetop-mostpoin tofeac hhemispheretoha v e asignican tlygreaterY-v aluethanthetopinnerpoin tofthein tra-cranialmask. GIVENYV alueofT op-MostP ointsofLeft/RightHemispheres: Y Left tb ;Y Righ t tb YV alueofCentroidofIntra-CranialRegion: Y ICR cb IF PreviousSliceisT emplate4 THEN IF max ( Y Left tb ;Y Righ t tb ) )Tj/T17 1 Tf50 0 TD(Y ICR t > 15 THEN Ma rkT emplate5 ELSE Ma rkT emplate4 T ransitionsfromT emplate1to4arelessob vioustomodel,butcanbedetected ifaslicehasasignican tdierencebet w eentherstmomen tofitsin tra-cranialmask andtherstmomen toftheheadgeneratedinSection4.2.1.InT emplate1slices, extra-cranialtissuesev enlysurroundthebrain,sothet w orstmomen tsarerelativ ely close.InT emplate4slices,ho w ev er,theey esbegintoappearandthefron tallobe beginstoshrink,causingtherstmomen tofthein tra-cranialregiontoshiftto w ards thebac kofthebrainanda w a yfromtherstmomen tofthehead.

PAGE 91

76 (a)OriginalMask (b)IsolatedCluster thatP assed (c)IsolatedCluster thatF ailed (d)RevisedMask (e)OriginalMask (f)ClusterwithF alse In tra-cranialRegion (g)Remo v edRegions (h)RevisedMask Figure18.ReningtheIn tra-CranialMask.Onceaslicehasbeenassignedatemplate,thein tra-cranialmaskisc hec k edagainstitsqualitativ emodel.Some slicesma yha v e\false"in tra-cranialclusters(b)orregions(g),suc hasthe ocularm usclescapturedb yapathologicalcluster.Spatialkno wledgecan detectandremo v esuc hregions/clusters,revisingthein tra-cranialmask (d)and(h).Ifarevisionissignican t,templateassignmen tisreconsidered. GIVENYV alueofCentroidofIntra-CranialRegion: Y ICR cb YV alueofCentroidofHead: Y h IF PreviousSliceisT emplate1 THEN IF j Y ICR cb )Tj/T17 1 Tf50 0 TD(Y h j > 7 THEN Ma rkT emplate4 ELSE Ma rkT emplate1 4.3.2 In tra-CranialMaskRenemen t Onceaslicehasbeenten tativ elyassignedatemplateandcorrespondingqualitativ emodel,kno wledgeconcerningthespatialarrangemen tofbraintissuesisused

PAGE 92

77 tolocatepossiblequalitativ einconsistenciesbet w eenthecurren tsliceandthetemplateten tativ elyassignedtoit.Mostsliceshadfew,ifan y ,inconsistencies.Insome slices,ho w ev er,especiallywheregadoliniuminfusedbloodhadenhancedextra-cranial tissues,\false"in tra-cranialregionsorclustersw erefoundafterbeingincorrectlycapturedinStageOne.Thepresenceof\false"in tra-cranialtissues,suc hastheey eand associatedtissuesinFigure18(e),couldcausedeviationsfromexpectedqualitativ e properties.Inmoreextremecases,templateassignmen tinSection4.3.1w asaected. Figure18(a),sho wsaT emplate4slicethatismislabeledasT emplate5duetothe presenceoftheey es.Inconsistenciesdonotnecessarilymeanthesliceisabnormal, onlythatextra-cranialtissuesw erepresen t,butma yaectpathologydetectionat laterstages. T emplateinconsistenciescanbedetectedandcorrectedusinganatomicalkno wledgeconcerningthespatialorganizationoftissueswithinthebrain. F orexample, whileaT emplate5sliceisdenedb ytheprojectionofthetemporallobesbey ondthe topinnerpoin t,thereisalimit(discussedbelo w)tothelengthofthatprojection, andregionsfoundbey ondthatlimitdonotbelongtothetemplatemodel. Using thisheuristic,theey einFigure18(e)canberemo v ed,resultinginFigure18(h).The o v erallprocessw orksasfollo ws: 1.F oreac hGroup2cluster, Cluster i ,isolateallpixelsbelongingto Cluster i that arecon tainedb ythein tra-cranialmask,creatingan\isolatedclusterimage." 2.F romeac hisolatedclusterimage,generatespatialregionsusinganeigh t-wise connectedcomponen tsoperation. 3.Comparethespatialcomponen tsof Cluster i withheuristicsthatdetectafalse in tra-cranialcluster.Ifafalsein tra-cranialclusterisdetected,remo v eitfrom thein tra-cranialmaskandplace Cluster i in toGroup1. Marktheslicefor templatereconsideration. 4.Compareeac hindividualspatialcomponen tfrom Cluster i withheuristicsthat detectfalsein tra-cranialregions.Ifaregionisfound,remo v eitfromthein tracranialmask.Iftheregionremo v edissignican t,marktheslicefortemplate reconsideration.

PAGE 93

78 Segmen tedImagewith F alseGroup2Cluster (X ,Y ) of Previous Slicetb tbRevised Threshold y Original Threshold y (b) Threshold y Ov er In tra-CranialMask (c)RevisedSegmen ted Image Figure19.Setting Threshold y .T odetectfalsein tra-cranialclusters,suc hastheey es in(a),aminim umdistancethreshold, Threshold y ,isused,assho wnin image(b).F oreac hGroup2clusterin(a),ifnosignican tcomponen t canbefound\behind"(lessthan) Threshold y ,suc hastheoneisolated indarkgra yin(b),thentheen tireclusterisremo v ed.Image(c)sho wsa revisedsliceafterremo v alofthefalsein tra-cranialclusterin(a). Examplesofisolatedclusterimagesaresho wninFigures18(b),(c),and(f).F alse in tra-cranialclustersareexaminedseparatelyandrstbecausetheyaresimplerto detectandma ysa v eprocessingb yiden tifyingtheen tireclusterinsteadofindividual spatialcomponen tscon tainedwithinit. Figure18(b)illustratesthemostusefulheuristicsofanormalin tra-cranial cluster,namelythatitsmostsignican tspatialcomponen tsofaclusteraredistributed throughoutthein tra-cranialregion.Incon trast,afalsein tra-cranialcluster,sho wnin Figure18(c),willha v eitsspatialcomponen tsconcen tratednearthefron tofthemask. Therefore,ifnospatialcomponen tof Cluster i canbefoundaminim umdistancea w a y fromthefron tofthemask/brain,thentheclusterisconsideredafalsein tra-cranial cluster. Theminim umdistancefromthefron tofthebrainisimplem en tedasav alue alongtheY-axis, Threshold y ,andaspatialregionpassesiftheY-v alueofitscentroidislessthan Threshold y .Theactualv alueof Threshold y issetbasedonthe in tra-cranialmaskofboththecurren tandpreviousslice,asw ellastheirrespectiv e templates. F ormostslices, Threshold y issettothetopinnerpoin tofthein tra-

PAGE 94

79 cranialmaskofthecurren tslice,(0 ;Y t ).Oftheslicesprocessedb ythissystem,the toppoin thasy ettobefoundatanareabelongingtoextra-cranialtissues,henceit isareliablelandmark. Ifatransitionbet w eenT emplate4to5isindicatedinSection4.3.1,ho w ev er, thenthepresenceoffalsein tra-cranialcluster(s)ma yha v ecausedtemplatemislabeling.T ouseasreliableav aluefor Threshold y aspossible,theboundingbo xofthe in tra-cranialmaskofthepreviousslice,whic hhasalreadybeencompletelyprocessed, isloadedandtheY-v alueofitstop-mostpoin t( X tb ;Y tb )isused. Once Threshold y hasbeendened,itismodiedwitha\modelfactor"that mo v es Threshold y eitherupordo wntheY-axis,dependinguponthetemplateofthe previousslice,tobetterisolateextra-cranialregionswhilepreservingalltruein tracranialregions.Ifthepreviousslicew asT emplate5, Threshold y ismo v edup w ard intheY-directionb y10%(m ultiply ing Threshold y b y1.1)toisolatethefron t-most tipsofthetemporallobeswhereextra-cranialregionsarelik elytobefound.Ifthe previousslicew asaT emplate1or4slice, Threshold y ismo v eddo wntheY-axisb y 10%(m ultiply ing Threshold y b y0.9). Anexampleofthisadjustmen tissho wnforthesliceinFigure19.AtransitionfromT emplate4to5w asindicatedasdenedinSection4.3.1,so Threshold y w asinitiallylocatedatthetop-mostpoin tofthein tra-cranialmaskoftheprevious slice.Thispoin tliesspatiallyjustoutsidethein tra-cranialmaskofthecurren tslice. Threshold y isreducedb y10%,inaccordancewiththemodelfactor.Thisisolatesthe extra-cranialcluster,sho wnindarkgra yinFigure19(b).Thisrevisionof Threshold y alsoisolatessometruein tra-cranialpixels,butbecausethesepixelsarespatiallyconnectedtootherpixelsbehindtherevised Threshold y ,theirresultingcen troidwill passtheheuristicandthepixelswillbepreserv ed. Oncethemodelfactordenedabo v ehasbeenset,therulefordetectingfalse in tra-cranialclusterscanbeimplem en te dasfollo ws:

PAGE 95

80 (0,Y )t Threshold y (a) Threshold y Ov er In tra-CranialMask Threshold y (b) Threshold y Ov er Segmen tedImage (c)RevisedMask Figure20.Remo vingF alseIn tra-CranialRegions. F oreac hGroup2cluster,ifan individualspatialcomponen tsisfoundinfron tof Threshold y ,thecomponen tisremo v ed. GIVENT opInnerP oint (0 ;Y ICRofCurren tSlice t ) \ModelF acto r"(MF),where: MF =0 : 9 ,ifthePreviousSliceisT emplate1o r4 MF =1 : 1 ,ifthePreviousSliceisT emplate5 Threshold y : IFPreviousSliceisT emplate4 ANDCurrentSliceisT entativelyT emplate5 THEN Threshold y = Y ICRofPreviousSlice tb MF ELSE Threshold y = Y ICRofCurren tSlice t MF F OREA CHGroup2 Cluster i IsolatePixelsBelongingto Cluster i fromtheIntra-cranialMask GenerateEight-wiseConnectedComponents Region i a F OREA CH Region i a GenerateCentroid ( X i a ;Y i a ) and Size i a (#ofpixels) IF :9 Region i a suchthat Y i a 50 pixels THENMa rk Cluster i as\F alseIntra-cranial"andRemove NoteIntra-cranialMaskfo rRevisionandT emplateRe-identication If Cluster i passestheruleslistedabo v e,theneac hofitsspatialcomponen tsistested tolocateextra-cranialregions,suc hastheonessho wninFigure18(g),remo vingan y individualcomponen twhosecen troidis\infron tof"(greaterthan) Threshold y ,as sho wninFigure20.Ifthepreviousslicew asT emplate1or4,anadditionalconstrain t

PAGE 96

81 isusedtoprotectsmallregionsatthefron tofthebrain,nearthein ter-hemispheri c ssure,b ylimitingthesystemtoconsiderationofonlyregionswithacen troidX-v alue morethan10pixelsa w a yfromtheY-axis,whereextra-cranialtissuesaremorelik ely tobefound.Duetotheprojectionofthetemporallobes,thisconstrain tisnotneeded ifthepreviousslicew asT emplate5. GIVENT op-mostP ointofLeftHemisphere Y Left tb T op-mostP ointofRightHemisphere Y Righ t tb Isolated Cluster i Image Eight-wiseConnectedComponents Region i a of Cluster i Centroid ( X i a ;Y i a ) and Size i a (#ofpixels)of Region i a IF PreviousSliceisT emplate1o r4 AND Y i a >Threshold y AND j X i a j 10 OR PreviousSliceisT emplate5 AND Y i a >Threshold y THENRemove Region i a F romMask IF Size i a > 75 pixels NoteIntra-cranialMaskfo rRevisionandT emplateRe-identication Ifaregionofsignican tsizeisremo v ed,thesliceismark edforresubmissiontothe heuristicsinSection4.3.1fortemplatereconsideration. Onceeac h Cluster i isprocessed,analc hec kforfalsein tra-cranialtissuesis madeb ysearc hingforclustersthatha v easharplyreduced(asdenedinEquation4.2) n um berofpixelsaftertherenemen tofthein tra-cranialmask,whic hsuggeststhat theclusterw ascomprisedmostlyofextra-cranialtissuesandshouldberemo v ed.F or eac hGroup2cluster,then um berofpixelscon tainedintherevisedin tra-cranialmask isdividedb ythen um berofpixelswithinthequadrangledescribedinSection4.2.1, formingtheratio: #PixelsinRenedIn tra-cranialMaskfor Cluster i #PixelsinQuadranglefor Cluster i (4.2)

PAGE 97

82 Iftheratioislessthan0 : 33, Cluster i ismark ed\F alseIn tra-Cranial"andplacedin Group1anditsremainingpixelsareremo v edfromthein tra-cranialmask.Onceall Group2clustersaretested,revisionofthein tra-cranialmask,ifneeded,isperformed usingthesamestepsdescribedinSection4.2.4whenextractingthein tra-cranial region. ArevisedmaskisthenresubmittedtotheT emplateiden ticationprocess describedinSection4.3.1.Otherwise,thesystemproceedstothenextsection. 4.3.3 T emplateV ericationwithV en tricleShapeInformation T ransitionsbet w eenT emplates1and4w ereinitiallydetectedb ytheheuristics describedinSection4.3.1b yasignican tdierencebet w eenthecen troidsofthehead andin tra-cranialregion.Accordingtokno wledgedescribedinSection2.3,ho w ev er, themostreliablepropert yisthebreakdo wnofthe\buttery"shapeofthewhite matterimmediatel ysurroundingthev en triclearea(V A).Thisiseasierandmore reliablethanexaminingtheCSFllingtheV AasCSFma ybecomprisedofoneor morespatialcomponen ts,whic hma yrequireamorecomplicatedqualitativ emodel. Instead,whitematterisisolatedanditsshapealongtheV Aisexamined. Ifthe \buttery"shapecanbefound,thesliceisconrmedasT emplate1. Otherwise, itdefaultstoT emplate4. OnceasliceislabeledT emplate4,templateordering describedinSection4.3.1dictatesthatnosuccessiv eslicecanbeT emplate1,making thisprocessingstepunnecessary 4.3.3.1 WhiteMatterSplitting Whitematteriskno wntobetheGroup2clusterwiththelo w estT2v alueforits cen troid.DuringinitialF CMsegmen tation,ho w ev er,o v er-clusteringcanoccasionally o v er-segmen twhitematterin tot w oneigh boringclusters.Whitematter\splitting"

PAGE 98

83 (a)AWhiteMatterClass (b)TheNeigh boringWhiteMatterClass (c)AWhiteMatterClass (d)TheNeigh boringGra yMatterClass (e)AWhiteMatterClass (f)TheNeigh boringGra yMatterClass Figure21.Tw oNeigh boringClassesinTw oSlices.

PAGE 99

84 ma yaectV AshapetestsinSection4.3.3.2andcausemisclassication.Therefore, an ysplitwhitematterclustersshouldbereco v ered. Figure21sho wsthelo w estandsecondlo w estT2v alueGroup2clustersrespectiv elyforthreeslices.Figures21(a)and(b)sho waslicewithwhitemattersplitin to t w oclasses,while(c),(d)and(e),(f)sho wawhitematterclassanditsneigh boring gra ymatterclass.At w o-lev elbinarydecisiontreew assuccessfullydev elopedb yLi in[70,71]fordetectingwhitemattersplittinginslicesin tersectingthev en tricles.A modiedv ersion,whic haddsadditionalkno wledgefromPD-w eigh tedspace,isimplemen tedinthissystem,andexaminesthepropertiesofthethreelo w estT2v alue Group2clusters,whic hrepresen t(inascendingT2order)kno wnwhitematter,a \candidate"clusterthatiseithergra ymatterorsplitwhitematter,andkno wngra y matter. Whitemattersplittingcanbedetectedintherstlev elofthedecisiontreeif thekno wnwhitematterandcandidateclusterha v esimilarpropertiesbasedeitheron bi-orthogonalthic kness(BT)oraclusterdensit ymeasure Density ,whic hisdened as: Density i = #Pixelsin Cluster i After5 5Erosion #Pixelsin Cluster i Before5 5Erosion (4.3) AsdescribedinSection2.7,aBTgiv esacoarseestimateofacomponen t's size.Hereitdetectsob viouscasesofwhitemattersplittingbecausetheBTofgra y matterisnev erm uc hlargerthantheBTofwhitematter[70,71].T ocomparecluster densities, Density i iscalculatedforboththekno wnwhitematterclusterandthe candidatecluster.Thet w odensit yv aluesarenormalizedtomak ethev alueofthe kno wnwhitematterclass( Density WM )notlessthan0 : 1.Thisisdoneb yrepeatedly m ultiplyi ngbothdensitiesb y10un til0 : 1 Density WM < 1. F orexample,if densitiesbeforescalingare Density WM =0 : 04and Density Candidate =0 : 008for anotherclass,thenafterscalingtheybecome0 : 4and0 : 08respectiv ely .Thedensit y

PAGE 100

85 ofgra ymatterislessthanwhitematter[70,71].Therefore,whitemattersplittingis detectedifbothnormalizeddensitiesaregreaterthan0 : 1. Whitemattersplittingcanalsobedetectedusingthedistributionofcluster cen tersinfeaturespace.Kno wledgedescribedinSection2.2andothersources[78, 32,106]indicatesthatprotondensit y(PD)isthebestpulsesequencefordiscriminatingbet w eenwhitematterandgra ymatter.Therefore,giv enthePDv aluesofthe cen troidsofakno wnwhitematterandgra ymattercluster,ifthecandidateclusteris splitwhitematter,thePDv alueofitsclustercen troidwillbeclosertothatofkno wn whitematter.Lik ewise,ifthecandidateclusterisgra ymatter,itwillliecloserin featurespacetothekno wngra ymattercluster.Thiscanbec hec k edb ycomparing thedistance(oftheclustercen troidsinPDspace)fromkno wnwhitemattertothe candidateclusteragainstthePDdistancefromthecandidateclustertokno wngra y matter. Thethreecomponen tsoftherstlev elofthedecisiontreecanbeimplem en ted inthefollo wingrule.Onlythoseslicesthatareten tativ elylabeledasha vingsplit whitematterproceedtothenextdecisionlev el. IF BT Candidate >BT WM +5 OR Density Candidate > 0 : 1 AND Density WM > 0 : 1 OR PD Candidate )Tj/T17 1 Tf49.9999 0 TD7 Tc(PD WM PD GM )Tj/T17 1 Tf50 0 TD7 Tc(PD Candidate THENT entativelyLabelCandidateClassasWhiteMatter andProceedtoDecisionLevelTw o. ELSETheCandidateClassisNotWhiteMatter Somecases,suc hasFigures21(e)and(f),ha v ewhiteandgra ymatterclusters withsimilarenoughfeaturestopasstherstlev elofthedecisiontree.Thesecond lev elusesanatomicalkno wledgetodiscriminatebet w eentheset w ocases. Figures22(a)and(b)sho wthein tra-cranialmasksimagesoft w oslicescorrespondingtoFigures21(a),(b)and(e),(f)respectiv ely .Bothpassedtherstlev el ofthedecisiontree.Anatomicalkno wledge,ho w ev er,indicatesthatgra ymattersur-

PAGE 101

86 (a)In tra-cranialMask (b)In tra-cranialMask (c)MergedImageof Figures21(a)and(b) (d)MergedImageof Figures21(e)and(f) (e)OriginalRing (f)OriginalRing (g)Subtracting (c)from(e) (h)Subtracting (d)from(f) Figure22.ShapeAnalysistoDetectWhiteMatterSplitting.

PAGE 102

87 roundswhitematterspatially .Therefore,itwilloccup ym uc hmoreoftheperiphery ofthein tra-cranialmask[70,71]. A\mergedclass"imageiscreatedb ymergingthekno wnwhitematterclass withthesuspectedsplitwhitematterclass,withallotherclassessettobac kground. Figure22(c)sho wsthemergedimagesofFigures21(a)and(b),whileFigure22(d) sho wsthemergedimagesofFigures21(e)and(f).Theperipheryofthein tra-cranial maskisappro ximatedb yapplyinga7 7erosionoperationtothein tra-cranialmask andsubtractingtheresultan timagefromtheoriginalmask.Figures22(e)and(f) sho wthein tra-cranial\rings"generatedfromthemasksinFigures21(a)and(b) respectiv ely .Anewringisgeneratedb ysubtractingthemergedclassimagefromthe originalring,assho wninFigures22(g)and(h),andthen um berofpixelsremaining (Figure22(g))isdividedb ythen um berofpixelsintheoriginalin tra-cranialring (Figure22(e)),formingaratio.Theratioisusedtodecideifaslicehaswhitematter splitting: IFRatio < 0 : 39 THENConrmCandidateClassasWhiteMatter ELSETheCandidateClassisNotWhiteMatter Theratiothresholdof0 : 39w asderiv edfromLi'seortsin[70,71]. Aslicewith whitemattersplittingwillt ypicallyha v earatiow ellabo v e0 : 39(0 : 88forthesliceof Figure22(g)).Sliceswithoutwhitemattersplittingwillha v esmallerratios(0 : 36for thesliceofFigure22(h)). 4.3.3.2 Chec kingtheShapeoftheV en tricleArea Onceallwhitematterclustersha v ebeeniden tied,theyaremergedtogether toformawhitematterimage,similartothatsho wninFigure22(c).A5 5median

PAGE 103

88 Right Frontal Horn (P75) Frontal MidPoint (P90) Left Frontal Horn (P105) Right Temporal Horn (P315) Temporal Mid-Point (P270) Left Temporal Horn (P225) Origin(a)P olygonAppro ximation ofV en tricleArea 0 L75 L90 L105 180 L235 L270 L315(b)EmitterLines 0 L75 L90 L105 180 L225 L270 L315 P315 P75(c)T emplate1Slice 0 L75 L90 L105 180 L225 L270 L315(d)T emplate4Slice Figure23.ExaminingtheV en tricleAreawithEmitter\Lines." lterisappliedtoremo v esmallpiecesofwhitemattertomak etheshapeofthewhite mattersurroundingthev en tricleseasiertoexamine. The\buttery"shapet ypicalofaT emplate1slicecanbebestmodeledb yusing apolygonappro ximationofthev en triclearea.Anexampleisgiv eninFigure23(a) andsho wsthesixpoin tsusedforthisappro ximation,four\horns"[102]andt w omidpoin ts,eac honeatanappro ximateangleintheCartesiansystemo v erlaidassho wn inSection4.2.1.F orexample,thefron tal(top)mid-poin tisat90degrees,alongthe v erticalaxis,whileitsneigh boring\horns"arefoundatappro ximately75and105 degrees. A t270degrees,thetemporal(bottom)mid-poin tcanbefound,withits neigh boringhornsat225and315degrees.Theanglesgiv endonotnecessarilymatc h thev en tricleareasofallslicesprecisely ,butpro videaw ork ablemodeltowhic ha

PAGE 104

89 slicebeingprocessedcanbecompared.Also,theimagesho wninFigure23(a)isonly at ypicalexampleandv ariationscanoccur,suc hassmallgapsinthewhitematter surroundingtheV A.Aqualitativ emodelisthereforeformedb yexaminingthemost reliablepropertiesacrosstrainingcasesofthesixpoin ts,bothindividuallyandthe relationshipsbet w eeneac h. 1.Thepoin tsatthefour\horns"arealw a ysenclosedb ywhitematter. 2.Thetemporalmid-poin tgenerallycon tactswhitematterandisclosertothe originthanitsneigh boring\horns." 3.Thefron talmid-poin tgenerallycon tactswhitematterandisclosertotheorigin thanitsneigh boring\horns." T oputthisqualitativ emodelinlo w-lev elterms,andtoallo wexaminationof theshapeofthewhitemattersurroundingthev en tricleareainthecurren tslice, \lines"(orprobes)areemittedout w ardfromthecen troidofthein tra-cranialmask, theoriginoftheCartesiansystem,atanglesof75,90,105,225,270,and315degrees, assho wninFigure23(b).Ifaline L i ,where i istheangleoftheemitterline,con tacts whitematter,theposition P i oftherstwhitematterpixelitmeetsisrecordedand theEuclideandistance, D i ,bet w een P i andtheoriginiscalculated.Ifaline L i does notreac hwhitematter, D i issetto )Tj/T20 1 Tf38.9999 0 TD(1. Onceallthe D i ha v ebeenrecorded,the\buttery"shapeofthev en triclearea ismodeledusingthefollo wingrule: IF D 75 0 AND D 105 0 AND D 225 0 AND D 315 0 AND D 270 0 AND D 270 max ( D 255 ;D 315 ) OR D 90 0 AND D 90 3 max ( D 75 ;D 105 ) THENLabelSliceT emplate1 ELSELabelSliceT emplate4 Theconditionthat D i 0requiresthattheemitterlineatangle i con tactwhite matter,enforcingtherstqualitativ eheuristiclistedabo v e. Thesecondheuristic isc hec k edb yv erifyingthatthedistancefromtheorigintothetemporalmid-poin t,

PAGE 105

90 whic hshouldcon tactwhitematter,isshorterthanthedistancetooneofitsneighboringhorns. The max ofthet w ohornsisusedtoallo wforthepossibilit ythat theheadissligh tlyrotated,causinganemitterlineateither225or315degreesto con tactthesideoftheV Ainsteadofreac hingtheactualtemporalhorn.This\prematurecon tact"w assligh tlymorecommonplaceforthefron talareaoftheV Adue tothepossiblepitc hofthehead(withthesubject'sheadtiltedsligh tlyupordo wn whileinsidetheMRcoil).T ocompensate,theheuristicforthefron talmid-poin tw as relaxedsligh tly Iftheslicepassestherule,itislabeledasT emplate1.Otherwise,thesliceis labeledasT emplate4.Anexampleofemitterlineso v erlaidonaT emplate1sliceis sho wninFigure23(c),while(b)sho wsemitterlinesforaT emplate4slice. 4.4 StageThree:P athologyDetection Onceanin tra-cranialmaskisgeneratedandaT emplatemodelassignedto it,theslicecanthenbetestedforenhancingpathology .Kno wledgeusedtodetect pathologywillcomefrombothanatomicalsourcesasw ellaspixeldistributionin featurespace.SliceswithdetectedpathologyareprocessedinChapter5fortumor segmen tation. ItshouldbenotedthatT emplate5Lslicesareprocessedonlyifthepreviousslicew asdeterminedtobeabnormalandonlyb ythemec hanismdescribedin Section4.4.3. Theabnormalshapesofthein tra-cranialmasksgeneratedb ythese slicesprev en tanatomicalkno wledgefrombeingconsisten tlyapplied.Morew orkon templatesfortheseslicesremains.

PAGE 106

91 4.4.1 Chec kingMaskSymmetry Accordingtoanatomicalkno wledge,anormalin tra-cranialmaskhasarough \symmetry"alongthev erticalaxis,wheresymmetryisdenedb yha vingappro ximatelythesamen um berofpixelsintheleftandrigh tbrainhemispheresinimage space,assho wninFigure24(a).Sliceswithsignican tincongruities,ho w ev er,suc h astheonesho wninFigure24(f),ha v eanincreasedlik elihoodofpathologybeing presen t.Theseincongruitiescanbedetectedandreco v eredb yapplyingareection operation,describedinSection2.7.2,tothein tra-cranialmaskandexaminingthe pixelsthatw erein troduced. Thereectionoperationisappliedaboutthev ertical axis,usingthecen troidofthein tra-cranialmaskastheorigin,andtheresultan t imageiscalledthe\symmetricalmask."Figures24(b)and(g)arethesymmetri cal masksgeneratedfromFigure24(a)and(f)respectiv ely Pixelsin troducedb ythereectionoperationareisolatedina\dierenceimage,"createdb ysubtractingtheoriginalin tra-cranialmaskfromthesymmetri cal mask,assho wninFigures24(c)and(h).Bothdierenceimageshadan um berof pixelsin troducedb ytheimagereection,whic hw aseithercausedb yasligh trotation oftheheadwithintheMRcoil(c),oractualdeformationsinthein tra-cranialmask (h),possiblyduetopathology .Inthenormalslice,Figure24(c),ho w ev er,allpixels in troducedliearoundtheperipheryofthesymmetri calmaskwithinarelativ elynarro wmargin.Incon trast,theabnormalslice,Figure24(h),hasaregionthatextends in tothein teriorofthebrainandism uc hmorecompactthanthoseinthenormalslice. Thisdierenceinshapesandlocationcanmodeledsothatpixelsin troducedb ythe reectionoperationduetothehead'srotationwithintheMRcoilcanberemo v ed. Anappro ximationoftheperipheryofthebrainiscreatedb yapplyinga9 9 erosionoperationonthesymmetri calmaskandsubtractingtheresultan timagefrom thesymmetric almask,assho wninFigures24(d)and(i).Mostoftheperipheralpixels

PAGE 107

92 Axis of Reflection(a) (b) (c) (d) (e) Axis of Reflection(f) (g) (h) (i) (j) Figure24.Chec kingIn tra-cranialMaskSymmetry .Images(a)-(e)arefromaslice withacceptablemasksymmetry ,whileanabnormalsliceissho wnin(f)(j).Areectionoperationisperformedonanin tra-cranialmask(a),(f) aboutthev erticalaxistogeneratea\symmetric al"mask(b),(g). The dierencebet w eenthesymmetric alandoriginalmaskissho wnin(c),(h). Aring(d),(i)appro ximatingtheperipheryofthesymmetric almaskremo v essmalldiscrepancies,sho wnin(e),(j).Only(j)hassignican tregions remaining. inadierenceimagecanberemo v edb ysubtractingtheresultingin tra-cranial\ring" fromthedierenceimage.A7 7medianlteristhenappliedtofurtherremo v e smallisolatedregions.Intheexampleslices,nearlyalloftheperipheralpixelsinthe normalslice(Figure24(e))w ereremo v ed,whiletheincongruit yinFigure24(j)w as preserv ed.Suc hincongruitiescanbedetectedb yapplyinganeigh t-wiseconnected componen tsoperation.Ifaspatialcomponen tofsignican tsize,175pixelsormore,is found,thesliceisconsideredtoha v e\abnormalmasksymme try"andthesymmetri cal maskisstoredandusedinplaceoftheoriginalin tra-cranialmaskfortheremaining pathologydetectionstages.Otherwise,theoriginalin tra-cranialmaskisused.

PAGE 108

93 4.4.2 Chec kingClusterSymmet ry Thesymmetryheuristicthatw asusedinSection4.4.1forthein tra-cranial maskisalsoapplicabletotheindividualin tra-cranialtissuescon tainedwithinit. Therefore,normalbraintissues/clusterswillha v egoodsymmetry(denedbelo w) whileclustercon tainingpathologywillha v epoorv erticalsymmetry Eac hGroup2 Cluster i isisolatedandtheset P i ( x;y ) ofallpixelsbelongingto Cluster i andcon tainedb ythein tra-cranialmaskiscreated.If( X c ;Y c )isthecen troid ofthein tra-cranialmask,thev alue Left i willcoun tthen um berofpixelsin P i ( x;y ) on theleftsideofthev erticalaxis,where P i x
PAGE 109

94 (a)In tra-Cranial Image (b)Abnormal Cluster (c)In tra-Cranial Image (d)Abnormal Cluster Figure25.SliceswithAbnormalClusterSymmetry anMRpixel'sv alueinT2andPDspaceincreases,itsT1v aluegenerallydecreases. Anexampleofthisprincipleissho wninT able4(a),whic hsho wst ypicalcluster cen terdistributionsforanormalslice.V erysmalluctuationscanbeseenbet w een neigh boringclusters,buttheo v eralltrendisconsisten t.Thistrendcanbeusedas aqualitativ emodelbecauseanabnormalsliceviolatesthisdistributionduetothe breakdo wnoftheblood-brainbarrier,whic hallo wsenhancingagen tstoen terthe in tra-cranialregionanddisturbtheexpectedpixeldistributions.F orexample,the sliceinT able4(b)hassignican tenhancingpathologypresen t,signiedb yacluster withhighcen troidv aluesinallthreefeatures. Arulethatdetectsabnormalitiesbasedonthedistributionsofthecluster cen tersw ouldw orkforcaseswithgrosspathology ,butitcouldalsoacceptslices wherepathologyissmallandmixedwithinagra ymatterorCSFcluster.P athology isinsteaddetectedatapixellev el.Sincekno wledgeindicatesthatpathologyoccupies thehighendofeac hfeaturespectrum,ifasignican tn um berofpixelswithhighT1, PD,andT2signalin tensit yv aluescanbelocated,thenenhancingpathologyhasbeen located.

PAGE 110

95 T able4.ClusterCen terDistributionsofNormalandAbnormalTissues.Giv ena t ypicalsetofclustercen troidsofin tra-cranialtissues,anormalslice,suc h astheexampleslicein(a),willha v earelativ elyconsisten trelationship bet w eentissuet ypesandclusterdistributioninfeaturespace,wherethe higheraPDandT2v alueatissuehas,thelo w eraT1v alueitwillha v e. Anabnormalslice(b)willviolatethistrend. (a)ClusterCen terDistributionofaSingleNormalSlice Tissue T1V alue PDV alue T2V alue WhiteMatter 392 782 315 WhiteMatter 391 830 356 Gra yMatter 360 901 390 Gra yMatter 362 973 438 Gra yMatter 334 991 516 CSF 270 1005 775 (b)ClusterCen terDistributionofaSingleAbnormalSlice WhiteMatter 388 777 295 WhiteMatter 379 859 337 Gra yMatter 370 935 392 Gra yMatter 363 998 502 P athology 786 1380 647 CSF&P athology 353 1092 730 4.4.3.1 MultispectralHistogramThresholding At ypicalslicewithenhancingpathologyissho wninFigure26.Thehistogram distributionsofthetumor,segmen tedb yaradiologist,w ereman uallyo v erlaidonthe histogramsofthein tra-cranialmask.Ineac hofthethreefeatures,thepathological pixels,sho wninblac k,arefoundatthe\high"endofthehistogram.Specicdimensionsarenotgiv eninFigure26,sincetheywillc hangefromslicetosliceandthe primaryconcernistherelativ elocationofenhancingpathologywithinthehistogram. Oneofthesimplestandmosteectiv emethodsofisolatingpixelsoneitherextreme ofanin tensit yhistogramisthroughthresholding.Thresholdingisasinglecon trast method,butsinceMRdataism ultispectral,thresholdscanbeseparatelyappliedin

PAGE 111

96 eac hofthethreefeatureimagesandtheresultscom bined.Enhancingpathologywill bedetectedifalargen um berofpixelssurviv ethethresholdsinallthreefeatures, withoneexceptiondiscussedinSection4.4.3.2. HistogramssimilartothoseinFigure26(c)through(e)arebuiltusingtheslice's in tra-cranialmask,orsymmetri calmaskifabnormalmasksymmetryw asdetected inSection4.4.1.Bothofthesemasksma ybeunreliableinT emplate5Lslices.Since aT emplate5Lsliceisprocessedonlyifthepreviousslicew asfoundtobeabnormal, ho w ev er,anassumptionismadethatpathologyis\spatially"adjacen tbet w eenslices andthetumorsegmen tationmask(describedinChapter5)oftheprevioussliceis usedasasubstitutein tra-cranialmask. Thehistogramthresholdsw eresetb yexaminingalltrainingslicesandobservingthat,ineac hfeatureimagesofaparticularslice,themajorit yofenhancing pathologyisfoundtotherigh t(ha vingagreatersignalin tensit y)ofthesignalintensit ybinha vingthegreatestn um berofpixels,thehistogram\peak."Basingthe thresholdonthehistogrampeakinsteadofaxedv aluegiv esthesystemmoreexibilit ywithindividualslices,butan um berofnon-tumorpixelscanalsobeseento therigh tofthehistogrampeaksinFigures26(c)through(e).Thesenon-tumorpixelsarenotofin terest.T oincreasethelik elihoodthatonlyenhancingpathologyis detected,thethresholdsintheT1andPDfeaturesareincreasedb y20%,whilethe T2thresholdisincreasedb y10%.Thesev aluesw ereempiricall yderiv edtomaximiz e theremo v alofallnormalbraintissues. Thera whistogrampeaksareusedinT emplate5Lslicesbecausetheirhistogramsaregeneratedusingthetumorsegmen tationoftheprevioussliceasthe in tra-cranialmask.Thisexcludesnearlyallnormalbraintissuepixels,whic hform thehistogrampeaksnormallyused. Ifthecurren tslicecon tainspathology ,then increasingthethresholdsma yremo v emostofthepathologyw eareattemptingto detect.

PAGE 112

97 (a)Ra wData (b)In tra-cranialMask T1-Weighted Value High T1 Low T1Pixel Count Intracranial Pixels "Ground Truth" TumorLow High(c)T1-w eigh tedHistogram PD-Weighted Value High PD Low PDPixel Count Intracranial Pixels "Ground Truth" TumorLowHigh(d)PD-w eigh tedHistogram T2-Weighted Value High T2 Low T2Pixel Count Intracranial Pixels "Ground Truth" TumorLow High(e)T2-W eigh tedHistogram Figure26.DetectingP athologyUsingHistogramDistributions.Enhancingpathology ,sho wninblac k,canbedetectedwithinasliceb ylocatingpixelswith highsignalin tensit yv aluesineac hofthethreefeaturespectrums.Note thatthepathologyisfoundtoha v eahighersignalin tensit ythanthe \peak"ineac hhistogram.

PAGE 113

98 (a) (b) (c) (d) Figure27.IsolatingP athologyThroughThresholding.ThresholdingthesliceinFigure26(a)inT1,PD,andT2-w eigh tedfeaturespaceproducesimages(a) through(c)respectiv ely .T akingthein tersectionofthethreeimagesresultsin(d),whic hcon tainsonlythosepixelswithhighsignalin tensitiesin allthreefeatures.Aslicewithasignican tn um berofpixelsin(d)most lik elycon tainsenhancingpathology Onceeac hthresholdoperationisappliedinitsrespectiv efeaturespectrum,a seriesofresultan timagesisproduced. Figures27(a)through(c)sho wtheresults ofthresholdapplicationinT1,PD,andT2-w eigh tedfeaturespacerespectiv ely .T o isolateonlythosepixelsthatsurviv edallthreethresholds,thein tersectionofthese threeimagesistak en.Anexampleoftheresultan timageissho wninFigure27(d)and issubsequen tlyreferredtoasthe\enhancingmask."T oremo v enoisypixelsandsmall piecesofmeningialtissue,a\wipe"operation(Section2.7.4)witha3 3windo wand athresholdof5.A3 3dilationisthenappliedtoconsolidateneigh boringpixels thatremainin tospatialregions,makingtheirsubsequen tanalysiseasier. 4.4.3.2 RegionAnalysisofAreasofEnhancemen t Giv enan\enhancingmask,"thedistributionheuristicdescribedinSection4.4.3 dictatesthatfew,ifan y ,pixelsshouldbepresen tinanormalslicewithanin tact blood-brainbarrier. T ov erifythis,eac h Cluster i fromtheinitialtenclassF CM segmen tationisisolatedandthen um berofpixelsbelongingto Cluster i con tained b ytheenhancingmaskarecoun ted.Ifno Cluster i canbefoundwith25ormore

PAGE 114

99 pixels(av alueusedduetothe3 3dilationabo v e),thesliceisconsideredtoha v e nosignican tabnormalities.Findingasuc h Cluster i ofsize > 25generallyindicates enhancingpathology ,anddoessoinT emplate5and5Lslices. InT emplate1and4slices,ho w ev er,theMRimagingplanema yin tersectwith bloodv esselsbeneaththetemporalv en tricles. Ingadoliniumenhancedcases,the partialv olumeeect,wheredieren ttissuet ypesareplacedin tothesameimaging v o xel,canmixgadoliniuminfusedbloodwithCSFinthetemporalv en tricles,resulting inpixelswithsignalin tensit yc haracteristicssimilartoenhancingpathology .Since T emplate5and5Lslicesdonotin tersectthev en tricles,thispossibilit ydoesnotapply tothem. Anexampleoftheseenhancingtemporalv en triclesissho wninFigure28,where (a)isthein tra-cranialregion(sho wingindividualclusterlabels)and(b)istheresultan tenhancingmask.Theenhancingtemporalv en tricles(referredtoafterw ardsas justthetemporalv en tricles)sho wninFigure28(b),ha v eapredictablestructurethat canbequalitativ elymodeledusinganatomicalkno wledge: 1.Thetemporalv en triclehornsha v egoodsymmetryalongthev erticalaxis. 2.Thetemporalv en triclesarefoundinthe\rear"ofthebrain,belo wthecen troid ofthein tra-cranialmask. 3.Thetemporalv en triclehornsmirroroneanotheratappro ximatelythesame anglefromthev erticalaxis. 4.Thetemporalv en triclesarealw a ysenclosedb ywhitematter. Althoughthetemporalv en triclessatisfyallfouroftheheuristicslisted,therstt w o areimplem en te dasseparaterulesbecauseenhancingmasksthatviolatethemcan beimm ediatel yclassiedasabnormal.Thelastt w oheuristicsarecom binedin toa singlerulethatsearc hesforthetemporalv en triclesandclassiesan ythingremaining aspathology

PAGE 115

100 Therstheuristic,concerningsymmetryalongthev erticalaxis,canbeillustratedb ycomparingenhancingmasksforpathology ,Figure27(d),andthetemporalv en tricles,Figure28(b). Themostnoticeabledierencebet w eenthet w ois theirrespectiv esymmetryalongthev erticalaxis,asdenedb y Symmetry i from Section4.4.2. Thetemporalv en triclesha v ev erygoodsymmetry ,whilepathology hasv erypoorsymmetry .Therefore,ifa Cluster i canbefoundwithasignican t n um berofpixelscon tainedintheenhancingmask, Size i ,withv erypoorsymmetry Symmetry i thenthenslicecanbelabeledabnormal.Theconceptisimplem e n tedas follo ws: IF 9 Cluster i suchthat: Size i > 50 pixelsAND Symmetry i < 0 : 10 THENLabeltheSliceAbno rmal Thesizerequiremen tisaddedtoprev en ttherulefromprematurelydeclaringaslice abnormalbasedonarelativ elysmalln um berofpixels. Althoughtherstheuristiccanbev alidatedataclusterlev el,theremaining heuristicsarebestmodeledandtestedusingregion-basedanalysis. Aneigh t-wise connectedcomponen tsoperationisappliedtotheenhancingmasktocreatean\enhancingcomponen ts"maskthatallo wseac hspatialregion, Region a ,tobeexamined separately .Thelocationoftheseenhancingcomponen tsinimagespaceisc hec k edb y usingemitterlines,in troducedinSection4.3.3.2.UsingthesameCartesianorien tationasinFigure23(b),Linesareemittedfromthecen troidofthein tra-cranialmask from0to360degrees,insteps45 = 4degreeusingthefollo wingalgorithm: j =0 WHILE j< 360 degrees,step 45 = 4 DO Emitline L Enhance j Emitline L Enhance j +1 Emitline L Enhance j )Tj/T15 1 Tf27 0 TD(1 ,whereif j =0 ,set j )Tj/T20 1 Tf50 0 TD[(1=359 Figure28(d)sho wstheemitterlinesatthemajorsteps j fortheenhancing componen timage.Additionallinesatangles j +1and j )Tj/T20 1 Tf51.0001 0 TD[(1areemittedtoensure

PAGE 116

101 (a) (b) (c) (d) (e) Figure28.EnhancingT emporalV en tricles. Gadoliniumcanoccasionallyenhance partsofthetemporalv en triclearea(b).Theseenhancingv en triclescanbe modeledandremo v edb ynotingthatthetemporalv en triclesareenclosed b ywhitematterneartherearofthebrain(c),andaresymmetricalong thev erticalaxis(e).Emitterlinesareusedin(d)and(e)tov erifythese rulesholdforanimage. con tactwithenhancingcomponen ts,possiblymissedatangle j eitherbecausethe componen tw astoosmallortheheadw assligh tlyrotated. Line L Enhance j returns theEuclideandistance, D Enhance j ,bet w eentheoriginandtherstenhancingpixel con tactedandtheiden tit yoftheenhancingcomponen t Region a j towhic hthepixel belongs(i.e., Region a w ascon tactedatangle j ).Ifnopixelofin terestw asencountered,line L Enhance j isremo v ed,lea vingonlythoselinesthatencoun teredenhancing componen tstobeconsidered. Giv enasetof\con tacts"withenhancingcomponen ts,theheuristicthatthe temporalv en triclesarelocatedattherearofthebraincanbeenforcedb yrequiring

PAGE 117

102 thatallenhancingcomponen ts, Region a j ,w ereencoun teredatanangle180
PAGE 118

103 GIVENAngle m Region m a D Enhance m D WM m Angle n Region n b D Enhance n D WM n WHERE a 6 = b AND n>m> 180 AND IF D Enhance m
PAGE 119

104 CHAPTER5 TUMORSEGMENT A TION Allslicesprocessedb ythetumorsegmen tationsystemha v epreviousbeenautomaticallyclassiedasabnormalb yoursystemandarekno wntocon tainglioblastomam ultiform etumorbasedonradiologistpathologyreports.T umorsegmen tationisan extensionofpathologydetection.Therefore,informationgeneratedduringpathology detectionstagesisreadilya v ailable. Asmen tionedinSection2.8,astrengthofthiskno wledge-basedsystemisits \coarse-to-ne"operation. Insteadofusingasingleprocessingstep,kno wledgeis appliediterativ ely ,easilyiden tiedtissuet ypesarelocatedandlabeledrst,then remo v edtoallo wa\focus"tobeplacedontheremainingpixels.Thetumorsegmentationsystemissimilarlydesigned,butwithasligh tv ariation.Byitsnature,tumor ism uc hmorediculttomodelincomparisontonormalbraintissues. Therefore, tumorisdenedheremoreb ywhatitisn'tthanwhatitis.Specically ,allpixels foundnottobeenhancingtumorareremo v ed,withan ythingthatremainsbeing labeledastumor. 5.1 StageOne:Reco v eringLostT umorintheIn tra-CranialMask ThepathologydetectionsystemsdescribedinSection2.8.2andChapter4for theupperandlo w erslicesrespectiv elyhadprocessingstepstoextractthebrainfrom therestoftheMRimage.Extraandin tra-cranialtissuesw eredistinguishedprimar-

PAGE 120

105 (a) (b) (c) (d) (e) Figure29.ReclaimingLostT umorPixels.(a)TheoriginalF CM-segmen tedimage; (b)pathologycapturedin\LostT umor"Group1clusters;(c)in tra-cranial maskusingonlyGroup2clusters;(d)maskafterincludingLostT umor clusterswithtumor;(e)maskafterextra-cranialregionsarecorrectly remo v ed. ilyb yseparatingtheclustersfromtheinitialF CMsegmen tationin tot w ogroups: Group2forbraintissueclusters(fromwhic hthein tra-cranialmaskiscreated),and Group1fortheremainingextra-cranialclusters.DuringF CMsegmen tation,ho wev er,enhancingtumorpixelscanoccasionallybeplacedin tooneormoreGroup1 clusterswithhighT1-w eigh tedcen troids,assho wninFigure29(b),andresultinan in tra-cranialmaskmissingpixelsofin terest,namelytumor.Anexampleofthiscan beseeninFigure29(c).Inordertosuccessfullysegmen ttumor,these\lost"tumor pixelsneedtobereco v ered.ThisisthegoalofStageOne,whic haddressesupperand lo w erslicesseparatelytobetterexploittheirrespectiv eanatomicalc haracteristics. 5.1.1 T umorReco v eryinUpperSlices Referringtoanatomicalkno wledge,tumorislocatedinthein tra-cranialregion whiletrueextra-cranialtissuesonlysurroundtheperipheryofthebrain.Therefore, aGroup1clustercon taininglosttumorpixelscanbeiden tiedifan um berofits pixelscanbefoundinthein tra-cranialregion.Ifasignican tn um beroftumorpixels w erelost,thein tra-cranialmaskma ybetoodistortedtousedirectly .Figures5(a) through(c)forT emplates1through3respectiv ely ,ho w ev er,sho wthatwhilethe

PAGE 121

106 in ternalstructuresofthebrainc hange,thebrainitselfmain tainsarelativ elyconsisten t\o v oid"shape.Thispropert yw asexploitedtoseparateextraandin tra-cranial clustersafterF CMsegmen tationthrougha\quadrangle"testlik etheonepresen ted inSection4.2.1andusedhere. Duringthequadrangletest,theQP(denedinSection4.2.1)ofeac hclusters w asrecorded.AGroup1 Cluster i isconsideredtopossiblyha v e\LostT umor"if morethan1%ofitspixelscanbefoundwithinthequadrangle,i.e. QP i 0 : 01. Thev alueof1%isusedtomaximiz ethereco v eryoflosttumorpixelsbecausetrue extra-cranialclusterswillha v ev eryfewpixelswithinthequadrangle,ifan yatall. PixelsbelongingtoLostT umorclusters(Figure29(b))aremergedwithpixels fromallGroup2clusters(Figure29(c))andsettoforeground(anon-zerov alue), withallotherpixelsintheimagesettobac kground(v alue=0). Thisproducesa \Reco v ered"in tra-cranialmasksimilartotheonesho wninFigure29(d). Sincea LostT umorclusterisaGroup1cluster,theReco v eredin tra-cranialmaskno walso con tainsareasofextra-cranialpixels.Figure29(d)sho wsthatinreco v eringthetumor inFigure29(b),pixelscorrespondingtotheey esandskin/fat/m uscle,whic harenot ofin terest,w erealsoadded. Smallregionsofextra-cranialpixelsareremo v edandseparationofthebrain fromconnectingtissuesisenhancedb yrst\in v erting"theReco v eredin tra-cranial mask,sothatforegroundpixelsoccup ythebac kgroundandvisa-v ersa,andapplying a5 5closingoperation. TheReco v eredin tra-cranialmaskisin v ertedagainand thebrainisextractedb yapplyinganeigh t-wiseconnectedcomponen tsoperation andk eepingonlythelargestforegroundcomponen t(thein tra-cranialmask).Finally \gaps"alongtheperipheryoftheReco v eredin tra-cranialmask,possiblytumornot reco v eredabo v e,arelledb yrstapplyinga15 15closing,thena3 3erosion operation.Thecen troidoftheresultan timageisthentak enandusedastheorigin forareectionoperation(Section2.7.2)aboutthev erticalaxis,producingthenal

PAGE 122

107 maskthatwillbepassedon toStageTw o.Anexampleofthenalmaskcanbeseen inFigure29(e). 5.1.2 T umorReco v eryinLo w erSlices Thesignican tc hangesintheshapeofthein tra-cranialregioninlo w erslices mak esusingQPv aluesunreliable. LostT umorclustersareinsteadreco v eredb y isolatingan yGroup1clusterwhocen troidv aluesaregreaterthanthecen troidof whitematterinT1,PD,andT2space,creatingan\EnhancingGroup1"mask.In thecasesofwhitemattersplitting,theclusterwiththehigherT1/lo w erT2v alue cen troidisusedasthecomparisoncluster.Theoriginalin tra-cranialmaskdescribed inSection4.3.2ismergedwiththeEnhancingGroup1masktocreateanimage similartothatsho wninFigure29(d). Eigh t-wiseconnectedcomponen tsisthen usedtoextractthelargestspatialcomponen t,thein tra-cranialregion,andcreatean \EnhancedIn tra-Cranial"mask. T oconsidersliceswhereatumorclusterw asplacedin toGroup2,buthadpixels lostduringrenemen tofthein tra-cranialmask,a\Group2"maskiscreatedloading theinitialF CMsegmen tationandsettingallGroup2clusterstoforegroundandall Group1clusterstobac kground.Aneigh t-wiseconnectedcomponen tsisapplied,but inadditiontothelargestspatialcomponen t,an yregionlessthan250pixels(alarger componen tiskno wntobepartoftheey es)isalsok ept.Thisimageisthenmerged withtheEnhancedIn tra-Cranialmaskandpixelsbelongingtomeningialtissuesare separatedfromtherestofthebrainwitha5 5erosionoperation.Thebrainisis extractedusinganothereigh t-wiseconnectedcomponen tsprocess,k eepingthelargest spatialregion,anditsoriginaldimensionsarerestoredwitha5 5dilationanda v erticalreectionoperationusingthemask'scen troidastheoriginforthev ertical axis.

PAGE 123

108 Itshouldbenotedthatthisprocessw orksforalllo w erslices,exceptT emplate 5L,wheretheseparationofthetemporallobesviolatestheabo v econstrain ts.Instead, aT emplate5Lslicewillusethemaskgeneratedinthissectionb ytheT emplate5 slicenearesttheshouldersasanappro ximationofitso wnmask. 5.2 StageTw o:MultispectralHistogramThresholding AfterStageOneiscompleted,therearethreeprimarytissuet ypesofin terest: pathology(whic hcanincludegadolinium-enhancedtumor,edema,andnecrosis),the brainparenc h yma(whiteandgra ymatter),andCSF,withtheultimategoalbeing tosegmen tthegadolinium-enhancedtumor. Eac hMRv o xelwillha v ea h T 1 ;PD;T 2 i locationin < 3 ,formingafeaturespacedistribution. Kno wledgedescribedinSection2.2andusedintheChapter4 algorithmincludesthefactthatpixelsbelongingtothesametissuet ypewillexhibit similarrelaxationbeha viors(T1andT2)andw atercon ten t(PD),theywillthen alsoha v eappro ximatelythesamelocationinfeaturespace[9].Figure30(a)sho ws thesignal-in tensit yimagesofat ypicalpathologicalslice,while(b)and(c)sho w projectionsinT1/PDandT2/PDspacerespectiv ely ,withappro ximatetissuelabels o v erlaid.Thereissomeo v erlapbet w eenclassesbecausethegraphsareprojections andalsoduetothepartial-v olumeeect. Kno wledgeusedinSection4.4.3fordetectingenhancingpathologyisalsoused foritsdelineation. T ypicalrelationshipsbet w eengadolinium-enhancedtumorand otherin tra-cranialtissuescanbeseeninFigure31,whic hconsistsofhistogramdistributionsforeac hofthethreefeatureimages.Thesehistogramsw ereexaminedand in terviewsw ereconductedwithexpertsconcerningthegeneralmak eupoftumorous tissueandthebeha viorofgadoliniumenhancemen tinthethreeMRspectrums.Lik e Figure26,specicdimensionsarenotsho wninFigure31,sincetheywillc hangefrom

PAGE 124

109 (a) T1-Weighted ValuePD-Weighted ValueC Pa Pa Pa T T THigh PD High T1(b) PD-Weighted ValueT2-Weighted ValueC Pa Pa Pa T THigh T2 High PD(c) Figure30.DistributionofIn tra-CranialPixels.(a)Ra wT1,PD,andT2-w eigh ted Data. Thedistributionofin tra-cranialpixelsaresho wnin(b)T1-PD and(c)PD-T2featurespace.C=CSF,P a=P arenc h ymalTissues,T= T umor slicetosliceandtheprimaryconcernistherelativ elocationofenhancingpathology withinthehistogram.Basedonthet ypicaltissuedistributions,suc hasthosesho wn inFigures30and31,andthegeneralMRprinciplesdiscussedinSection2.2,asetof heuristicsw ereextractedthatcouldbeincludedinthesystem'skno wledgebase. 1.Gadolinium-enhancedtumorpixelsoccup ythehigherendoftheT1spectrum. 2.Gadolinium-enhancedtumorpixelsoccup ythehigherendofthePDspectrum, thoughnotwiththedegreeofseparationfoundinT1space[49]. 3.Gadolinium-enhancedtumorpixelsw eregenerallyfoundto w ardsthe\middle" oftheT2spectrum,makingreliablesegmen tationbasedonT2v aluesdicult. 4.Sliceswithgreaterenhancemen thadbetterseparationbet w eentumorandnontumorpixels,whilelessenhancemen tresultedinmoreo v erlapbet w eentissue t ypes.

PAGE 125

110 (a)Ra wData T1-Weighted Value High T1 Low T1Pixel Count Intracranial Pixels "Ground Truth" Tumor(b)T1-w eigh tedHistogramPD-Weighted Value High PD Low PDPixel Count Intracranial Pixels "Ground Truth" Tumor(c)PD-w eigh tedHistogram T2-Weighted Value High T2 Low T2Pixel Count Intracranial Pixels "Ground Truth" Tumor(d)T2-W eigh tedHistogram Figure31.HistogramDistributionsforT umorandtheIn tra-CranialRegion.Solid blac klinesindicatesthresholdsinT1andPD-w eigh tedspace,whic hw ere basedonthehistogram\peaks."

PAGE 126

111 Theseheuristicsareexpressedingeneraltermstogiv ethesystemmoreexibilit ywhenprocessingsliceswithsligh tlydieren tscanningprotocols. Thisisin con trasttoeortsthattuneimagingparameters,suc has[112],andthereforeha v e parameterdependen ttrainingsets,whic hlimitstheirapplicationtosliceswiththe sameparameters. Uptothispoin t,thesystemhasusedaparadigmwhereadatasetw ould bepartitionedwithanunsupervisedclusteringalgorithm,witheac hpartitionbeing labeledaccordingtoruleswithinthekno wledgebase.Thisparadigmisexpandableto aniterativ ev ersionwhere,giv enaninitialsegmen tationgeneratedb yo v er-clustering, easilyiden tiableobjects/clustersw erelabeledandremo v edandtheremainingdata w as\reclustered"toallo wmoresubtletrendsbet w eenobjectsofin teresttobecome clearforsubsequen trules.Called\kno wledge-basedreclustering,"thisapproac his implem en te dinthissystem,asdetailedin[20,18],forreningtheboundariesbet w een gra ymatterandCSFinnormalslices. W orkb ythisauthorin[22]sho w edthe paradigmw asalsoapplicabletootherdomains,suc hassatelliteimaging. T ousekno wledge-basedreclusteringfortumorsegmen tation,oncethein tracranialmaskw ascreated(th usdiscardingallextra-cranialpixels),reclusteringw ould beappliedtothosepixelscon tainedinthemasktoseparatenormalbraintissuesfrom pathology .Allnormaltissuesw ouldthenbediscardedandtheremainingpixelsw ould bereclusteredtoseparateenhancingtumorfromnon-tumorpathology .Whilew orking w ellforcaseswithmoredistincttumorboundaries,asreportedb ythisauthorin[5 19],whentumorswithmorediuseboundariesw ereconsidered,thereclusteringstep w ouldarriv eatpartitionsthat,whilenot\meaningless"[126,123],ga v epoorresults whenlabelsw ere(man ually)assignedtov ariousclusters.F orexample,asignican t n um beroftumorouspixelsw erebeingmisplacedin toclassesofnormalbraintissue, suc haswhitematter. A ttemptstosolv ethisproblemin v olv edgreaterlev elsof o v er-clusteringandan um berofF CMv ariations,includingthesemi-supervisedF CM

PAGE 127

11260 61 62 63 64 65 66 67 68 567891011121314Jm ValueFCM Iteration (a)11 12 13 14 15 16 567891011121314Number of ErrorsFCM Iteration (b) Figure32. J m V ersustheNum berofErrorsfortheIrisSet. Theprogressofthe optimizationfunction J m o v erF CMiterationsissho wnin(a),whilethe n um beroferrorsateac hiterationissho wnin(b). algorithm(ssF CM)[4](usedtoin troducetrainingdataautomaticallyselectedb ythe kno wledgebase),Bensaid'sv alidit y-guidedclustering(V GC)[3,123],andGustafson andKessel'sfuzzyco v ariancemethod[43]. Noneoftheseapproac hescompletely solv edtheproblem. Indev elopingtheV GCmethod,Bensaid[3 ]iden tiedan um beroflimitationsin unsupervisedclustering.Themostrelev an tlimitationisthefactthatanunsupervised clusteringalgorithmoptimizesafunctionthatma ynotbeagoodestimatorof\true" classicationqualit y .Mostc-meansclusteringalgorithmsoptimizeawithingroups sumofsquarederrorsfunction, J m .Afuzzyv ersionisgiv eninSection2.5.Man y \realw orld"datasets,ho w ev er,ha v eoptimalpartitions,determinedb yexpertsin thateld,thatdonotnecessarilycorrespondtotheoptimalv alueoftheobjectiv e function.AnexampleofthiscanbeseeninFigure32fortheIrisdataset.Comparing thev alueof J m withthen um beroferrors,itcanbeseenthatapartitionwithahigher J m hasfew ererrorsthanthe\optimal"partitionaccordingto J m F orthespecicproblemoftumorsegmen tation,analysisofkno wledgeconcerningtissuecon trastsinspecicpulsesequencesandthepixeldistributionsinfeature spacerev ealedthattheT2-w eigh tedfeaturepla y edatmostasmallroleindeterminingthebestboundariesbet w eentumorandnon-tumorpathology .Infact,gadoliniumsligh tly reduces theT2-w eigh tedsignalin tensit yofaectedtissues[49].While

PAGE 128

113 exhibitingsligh tlybetterdecisionboundaries,experimen tswithclusteringusinga reducedn um beroffeatures(T1andPDonly)sho w edthatF CMstillhadatendency toplacetumorpixelsin toclusterscon tainingprimarilynormaltissuesinsliceswith diuseboundaries. AnalysisofthehistogramsgeneratedinFigure31toexamine therelationshipbet w eenenhancingtumorandotherin tra-cranialtissues,ho w ev er, rev ealedthatthresholdingcouldpro videasimple,fast,y eteectiv e,mec hanismfor grossseparationoftumorfromnon-tumorpixelsandimpleme n tationoftheheuristics listedabo v e.Thisnotionisv erysimilartothepathologydetectionmec hanismdescribedinSection4.4.3.Infact,theuseofm ultispectralhistogramthresholdingw as rstappliedfortumorsegmen tation,thenlateradaptedtothemoregeneralproblem ofpathologydetection. Itshouldbestressed,ho w ev er,thattheuseofhistogramthresholdingdoesnot, andshouldnot,suggestthatclusteringisincapableofarrivingatsatisfactoryresults. Eortsb ythisauthorin[5,19]ha v esho wnthatkno wledge-based(re)clusteringis arobustparadigm.Butthehistogrambasedmethodisasimplerandfasterw a yof exploitingtheheuristicslistedabo v etoobtainusefulresults. Comparisonsofthe histogram-basedmethodwithssF CMsegmen tations(whic husedman uallyselected trainingexamples)aremadeinT able17inChapter6andsho wthatssF CMperformed w ellwithtumorsthathaddistinctboundaries,suc hasP atien t1,butsignican tly underestimatedintumorcasesthathadmorediuseboundaries,suc hasP atien t2. AsnotedinSection4.4.3,intheT1andPDspectrums,thev astmajorit yof enhancingtumorpixelscanbefoundtotherigh t(ha vingagreatersignalin tensit y v alue)ofthehistogram\peak,"whic hserv esasaneectiv ethresholdthatw orks acrossslices,ev enthosewithv aryingdegreesofgadoliniumenhancemen t. While theT2featurew asanadditionalsafeguardforensuringonlyenhancingpathology surviv edinSection4.4.3,asnotedabo v e,gadolinium'stendencytoreducetheT2 signalin tensit yofaectedtissuesmadeitun usableforthemorepreciseproblemof

PAGE 129

114 (a) (b) (c) (d) Figure33.MultispectralHistogramThresholdingofFigure31. (a)T1-w eigh ted thresholding;(b)PD-w eigh tedthresholding;(c)In tersectionof(a)and (b);(d)Groundtruth. tumorsegmen tationandw asexcludedfromthisprocess. Anexampleissho wnin Figure31. Unlik einSection4.4.3,wherethehistogramthresholdsw ereadjusted tomaximiz etheremo v alofallnon-tumorpixels,thetumorsegmen tationsystemis designedtominim izetumorlosswhenremo vingnon-tumorpixels. Therefore,the thresholdsherearenotadjustedastheyw ereinSection4.4.3.Also,whilethepeak methodw orksw ellforallslicesT emplate1through5,thehistogramdistributions forT emplate5Lslicesw eresligh tlydieren tduetothefactthattheyareusingan in tra-cranialmaskfromahigherslice.ThisdistortsthePDhistogramdistribution sho wninFigure31,butthiscanbecompensatedb yusing Peak PD 1 2 asthePD threshold,whiletheT1thresholdma ybesetnormally Figure33(a)and(b)sho wtheresultsofapplyingtheT1andPDhistogram thresholdsinFigure31(b)and(c),allo wingdirectcomparisonstobemadebet w een thehistogramsandtheresultsoftheirthresholds.Inbothofthesethresholdedimages asignican tn um berofnon-tumorpixelsha v ebeenremo v ed,thoughsomenon-tumor pixelsremainineac hthresholdedimage.Sincetumoriskno wntoha v ehighsignal in tensitiesinbothT1andPDspace,ho w ev er,thein tersectionoftheset w oimageswill furtherremo v enon-tumorouspixelswhilepreservingtumor.Anexampleissho wnin Figure33(c).

PAGE 130

115 Low PD High PD Low T1 High T1Number of PixelsLow High(a)2D-HistogramProjection.Low PD Low T1High T1High PD(b)ScatterPlot. Low T1High T1Low PD High PD (c)AfterScreening. (d)InitialT umor (e)T umorAfter Densit yScreening (f)GroundT ruth. Figure34.Densit yScreeningInitialT umorSegmen tationF romFigure33(c). 5.3 StageThree:\Densit yScreening"inF eatureSpace Them ultispectralhistogramthresholdingprocessinStageTw opro videsagood initialtumorsegmen tation,suc hastheonesho wninFigure33(c). Comparingit withthegroundtruthimageFigure33(d),an um berofpixelsintheinitialtumor segmen tationarenotfoundinthegroundtruthimageandshouldberemo v ed.A tthis stage,ho w ev er,additionalthresholdingisdiculttoapplywithpossiblyremo ving tumorpixels. AssummarizedinSection2.2,pixelsbelongingtothesametissuet ypewill ha v esimilarsignalin tensitiesinthethreefeaturespectrums.Becausenormaltissue t ypesha v eamoreorlessuniformcellularmak eup[78,32,106],theirdistributionin

PAGE 131

116 featurespacewillberelativ elyconcen trated[9].T umor,ho w ev er,canha v esignican tv ariance,dependingonthelocaldegreesofgadolinium-enhancemen tandtissue inhomogeneit ywithinthetumorduetothepresenceofedema,necrosis,andpossiblysomeparenc h ymalcellscapturedb ythepartial-v olumeeect.Figures30(b) and(c)sho wthedieren t\spreads"ofnormalandtumorpixels.Pixelsbelonging toparenc h ymaltissuesandCSFaregroupedmoredenselyb yin tensit y ,whilepixels belongingtotumoraremorewidelydistributed. This\densit y"c haracteristiccanbeusedtoremo v enon-tumorpixelswithout aectingthepresenceoftumorpixels.Called\densit yscreening,"theprocessuses a3-dimensionalhistogramforallpixelsremaininginthetumorsegmen tationimage afterStageTw o. Thehistogramarra yhasa T 1 range PD range T 2 range sizeof128 128 128in tensit ybins.Themaxim umandminim umsignalin tensit y v aluesofeac hfeatureintheinitialtumorsegmen tationarefoundandquan tizedin to thehistogramarra y(i.e.,theminim umT1in tensit yv alueoccupiesT1Bin1,the maxim umT1in tensit yv alueoccupiesT1Bin128,withotherv alues\quan tized" inbet w een). Thequan tizationw asdonefort w oreasons. First,sizesofathreedimensionalhistogramquic klybecameprohibitiv elylargetostoreandmanipulate. Ev ena256 3 histogramhasnearly17millionelemen ts.Secondly ,lev elsofquan tization canmak ethe\dense"natureofnormalpixelsclearwhilestilllea vingtumorpixels relativ elyspreadout.F orthe12-bitdatastudiedhere,afterhistogramthresholding, sliceshadarangeofappro ximately800in tensit yv aluesineac hfeature.Theactual v alueof128w asempiricallyselectedafterafactorof256,astandardv alueinthe n um berofin tensitiesingra yscaleimages,w asfoundtobeun wieldytouseanddid not\enhance"thedensit yofnormalpixels.Using64binsineac hfeatureblurred theseparationoftumorandnon-tumorpixelsintrainingsliceswherethetumor boundaryw asnotasw elldened. V aluessimilarto128,suc has120or140,are unlik elytosignican tlyc hangethe\quan tization"eect,norshouldbrainsizesince

PAGE 132

117 thehistogramsarebasedonthefeaturev aluesofsuspectedtumor.Thehistograms andscatterplotssho wninFigure34w erecreatedusing128bins.Specicdimensions arenotsho wninFigure34,sincetheywillc hangefromslicetosliceandtheprimary concernisthelocationandrelativ edensit yofnon-tumorpixelsincomparisonto tumorpixels. F romthe3Dhistogram,three2Dprojectionsaregenerated:T1/PD,T1/T2, andPD/T2.Anexample2Dprojectionissho wninFigure34(a),generatedfromthe slicesho wninFigure33(c).Non-tumorpixelsarelocatedinthelo w estT1/PDcorner (consisten twithkno wledgeinfeaturespace)andaresho wntoha v ethehighestpeaks, meaningthosebinsha v ethemostpixelswithinthem.Acorrespondingscatterplotis sho wninFigure34(b).Similarpropertiescanbefoundintheothert w oprojections. T oremo v etheseareasofdensenon-tumorpixels,thehighestpeakineac hprojectionisfoundanddesignatedasthestartingpoin tforaregiongro wingprocess[55] that\clears"an yneigh boringbinwhosecardinalit y(n um berofpixelsinthatbin)is greaterthanasetthreshold(T1/PD=3,T1/T2=4,PD/T2=3).Thiswillresultina newscatterplotsimilartothatsho wninFigure34(c).Apixelisremo v edfromthe tumorsegmen tationifitcorrespondstoabinthatw as\cleared"inan yofthethree feature-domainprojections.Figures34(d)and(e)arethetumorsegmen tationbefore andaftertheen tiredensit yscreeningprocessiscompleted.Notethattheresulting imageisclosertogroundtruth. Thethresholdsusedw eredeterminedb ytakingthegroundtruthtumorof eac htrainingsliceandcreatinga3Dhistogram,including2Dprojections,based onthedimensionsoftheslice'sinitialtumorsegmen tation. Inotherw ords,giv en a3Dhistogramofaninitialtumorsegmen tation,allpixelsnotinthegroundtruth imagew ereremo v ed,lea vingonlytumorbehindwithoutc hangingthedimensionsand quan tizationlev elsofthehistogram. Therespectiv e2Dprojectionsofalltraining slicesw ereexamined. Itw asfoundthatthesmallestbincardinalit yborderinga

PAGE 133

118 binoccupiedb ykno wnnon-tumorpixelsmadeanaccuratethresholdforthegiv en projection. 5.4 StageF our:RegionAnalysis Uptothispoin t,kno wledgefortumorsegmen tationhasbeenappliedprimarily toapixel'sindividualpropertiesinfeaturespace. StageF our,ho w ev er,in tegrates spatialkno wledgeb yconsideringpixelsataregionorcomponen tlev el. Dened inSection2.7.3,aconnectedcomponen tsoperationisappliedtotherenedtumor segmen tationgeneratedb yStageThree,allo wingeac hregiontobetestedseparately forthepresenceoftumor.Figure35(b)sho wsarenedtumorsegmen tationmask, whic hcon tainsan um berofspatiallydisjoin tareas. Thegroundtruthtumor(c) sho wsthatonlyoneregionactuallycon tainstumor. Therefore,decisionsm ustbe maderegardingwhic hregionscon taintumorandwhic honesdonot. Whilemostglioblastoma-m ultiform etumorcasesha v eonlyonetumorousspatiallycompactregionthathasthehighestmeanT1v alue,insomecases,thetumor hasgro wnsuc hthatithasbranc hedin tobothhemispheresofthebrain,causingthe tumortoappeardisjoin tinsomeslices,orithasfragmen tedasaresultoftreatmen t. Also,dieren ttumorregionsdonotenhanceequally .Th us,casescanrangefroma singlew ell-enhancingtumortoafragmen tedtumorwithdieren tlev elsofenhancemen t.Incomparison,themak eupofnon-tumorregionsisgenerallymoreconsisten t thanintumorousregions.Therefore,thekno wledgebaseisdesignedtofacilitateremo v alofnon-tumorregionsbecausetheircompositioncanbemorereliablymodeled anddetected.

PAGE 134

119 (a) (b) (c) Figure35.UsingRegionsforT umorSegmen tation.Afterprocessingthein tra-cranial mask(a),(b)isarenedtumorsegmen tation.Onlyoneregion,assho wn intheground-truthimage(c)isactualtumor.Regionanalysisdiscriminatesbet w eentumorousandnon-tumorousregions. 5.4.1 Remo vingExtra-CranialRegions T able2sho wsthatbothgadolinium-enhancedtumorandextra-cranialtissues tissuesthatreceiv egadoliniuminfusedblood,suc hasm uscleormeningialtissues immedi atelysurroundingthebrain,willha v eabrigh tT1signalin tensit y .Thepresenceofsuc hextra-cranialpixelscanin terferewiththekno wledgebase'sassumption inSection5.4.2thatregionswiththehighestT1v aluearemostlik elytumor.Th us, theirremo v alisimportan ttoallo wtheheuristicsinSection5.4.2tobeconsisten tly applied. 5.4.1.1 ProcessingLo w erSlices Inthelo w erslices,theextra-cranialregionsmostoftenrein troducedduringtumorreco v eryaretheocularm usclesandnerv es.Anexampleissho wninFigure36(a). Thesestructuresareeasilyrecognizedb ythefactthattheyarepairsofcomponen ts ofappro ximatelythesamesize(n um berofpixels)andmirroroneanotheralongthe

PAGE 135

120 (a) (b) (c) Figure36.Remo vingtheEy es.Reco v eringlosttumorpixelscanoccasionallyreintroduceextra-cranialstructuressuc hastheey es(a).Thesestructurescan bemodeledandiden tied(b),allo wingtheirremo v al(c). v erticalaxisatthefron tofthehead.Sincetheyare\normal"structures,theyma y bequalitativ elymodeledandremo v ed. Thesizeheuristiccanbetestedb yrequiringthatthen um berofpixelsineac h regiondierb ynomorethan25%.Componen tswillmirroroneanotheriftheyha v e appro ximatelythesamecen troidY-v alueandX-distancefromthev erticalaxis.The \fron t"oftheheadisdeterminedb yloadingtheoriginalin tra-cranialmask(before tumorreco v ery)andttingaboundingbo xaroundit. Thisreturnsthecen troid ofthemask( X ICR c ;Y ICR c ),usedastheoriginforaCartesianaxis,andthetop-most andbottommostpoin ts,fromwhic hthe\length"alongthey-axisofthein tra-cranial region Length ICR y iscalculatedb ytakingtheabsolutedierenceoftheYv aluesofthe top-mostandbottommostpoin tsoftheboundingbo x.Aregionisconsideredtobe atthefron toftheheadiftheY-v alueofitscen troidisnomorethan Length ICR y 25% belo wthetop-mostpoin t, Front ICR y .Thefollo wingruleimpleme n tstheheuristics:

PAGE 136

121 GIVENF rontThreshold Front ICR y Region a withCentroid ( X a c ;Y a c ) and Size a Region b withCentroid ( X b c ;Y b c ) and Size b IF min ( Size a ;Size b ) max ( Size a ;Size b ) 0 : 25 AND Y a c >Front ICR y AND Y b c >Front ICR y AND j Y a c )Tj/T17 1 Tf49.9999 0 TD(Y b c j < 5 AND j X a c j)-1000(j X b c j < 5 THENRemove Region a and Region b Tw oothertestsarealsousedtoremo v eextra-cranialregionsinlo w erslices. Therstremo v esan yregionwhosecen troidliesoutsidetheboundingbo xofthe originalin tra-cranialmask.Th us,acomponen tthatw astotheleftoftheleft-most poin tw ouldberemo v ed,asw ouldacomponen tthatw asabo v ethetop-mostpoin t, etc. Thesecondtestsearc hesforcomponen tsintherenedtumormaskthatha v e littleornocon tactwiththeoriginalin tra-cranialmask(beforetumorreco v eryin StageOne). Thisreectsthefactthattumorreco v eryw asprimarilydesignedfor pixelsacciden tallyclusteredin toanextra-cranialclusterandan ylargecomponen t foundintherenedtumormaskwithlittleornocon tactinthein tra-cranialmask isv erylik elytobeextra-cranialtissue.Thisisdoneb yisolatingeac hspatialregion intherenedtumormask,in tersectingitwiththeoriginal-mask,andcomparingthe n um berofpixelsthatsurviv ewiththeoriginaln um berofpixelsintheregion.Giv en Region a andthen um berofpixels Size a ,aratio Contact a canbedenedasfollo ws: Contact a = #Pixels(in Region a \ In tra-CranialMask) Size a (5.1) andv eriedwiththefollo wingrule(using Size a tolimittheruletoregionswithinan expectedsizerange,basedonobserv ationsmadeonthetrainingslices,ofextra-cranial tissues):

PAGE 137

122 GIVEN Region a with Size a IF Contact a =0 OR 300 Size a 800 AND Contact a < 0 : 075 THENRemove Region a 5.4.1.2 Remo vingMeningialRegionsinAllSlices Onceallextra-cranialregionsspecictothelo w erslicesha v ebeenremo v ed, additionalprocessingbeginstoremo v emeningialregionsfoundinallslices,suc has theduraorpiamatter.Anatomicalkno wledgeisusedagainb ynotingthatmeningial tissuesarethinmem branessurroundingthebrainandliewithinarelativ elynarro w margin.Therefore,an yregionsintherenedtumorimagethatha v easimilarplacemen tandshapew ouldbeexpectedtobemeningialtissues. Anexampleofsuc h meningialpixelsissho wninFigure29(e). Figure37sho wsthatanappro ximationofthebrainperipherycanbeusedto detectmeningialtissues.Theun usualshapeofthein tra-cranialmaskinFigure37(a) isduetopriorresectionsurgery .Thebrainperipheryisappro ximatedb yapplyinga 7 7erosionoperationtothein tra-cranialmaskandsubtractingtheresultan timage fromtheoriginalmask,assho wninFigures37(a)through(c).Eac hcomponen tin therenedtumormaskisin tersectedwiththein tra-cranialborder.An ycomponen t ha vingmorethan50%ofitspixelscon tainedinthein tra-cranialperipheryismark ed asmeningialtissueandremo v edfromfurtherprocessing.Figure37(d)sho wsatumorsegmen tationin tersectedwiththein tra-cranialperipheryfromFigure37(c).In Figure37(e),thepixelsthatwillberemo v edb ythisoperationaresho wnandthey areindeedmeningialpixels.

PAGE 138

123 (a) (b) (c) (d) (e) Figure37.Remo vingMeningialPixels.A\ring"appro ximatingthebrainperiphery iscreatedb yapplyinga7 7erosionoperationtothein tra-cranialmask (a),resultinginimage(b). Subtracting(b)from(a),createsa\ring", sho wnin(c).Byo v erla yingthis\ring"on toatumorsegmen tation(d), smallregionsofmeningialtissues(e)canbedetectedandremo v ed.The un usualshapeofthein tra-cranialregionisduetopriorresectionsurgery 5.4.2 Remo vingIn tra-CranialNon-T umorRegions Oncean yextra-cranialregionsha v ebeenremo v ed,thekno wledgebaseisappliedtodiscriminateregionswithtumorfromregionswithouttumorbasedonstatisticalinformationabouttheregion.Aregionsmean,standarddeviation,andsk ewness inT1,PD,andT2featurespacerespectiv elyareusedasfeatures.F orexample,the meanT1v alueof Region a u a T 1 ,w ouldbedenedas:

PAGE 139

124 u a T 1 = 1 N N X i =1 T 1( x i ) (5.2) where T 1( x i )istheT1v alueofpixel x i con tainedin Region a .Thev alue N represen ts then um berofpixelsin Region a andislaterreferredtoas Size a Thestandard deviationof Region a ( a T 1 )isdenedas: a T 1 ( x 1 :::x N )= v u u t 1 N )Tj/T20 1 Tf50 0 TD(1 N X i =1 ( T 1( x i ) )Tj/T17 1 Tf49 0 TD(u a T 1 ) 2 (5.3) anditssk ewnessis: Skewness a T 1 ( x 1 :::x N )= 1 N N X i =1 T 1( x i ) )Tj/T17 1 Tf50 0 TD(u a T 1 a T 1 3 (5.4) Theconceptbehindthisapproac histhatthetrendsandc haracteristicsdescribedatapixellev elinSections2.2and5.2arealsoapplicableonaregionlev el. F orexample,sinceindividualpixelswithahigherT1v aluearemorelik elytocon tain tumorthanpixelswithalo w erT1v alue,oneexpectsregionswithahigherT1mean aremorelik elytocon taintumorthanregionswithalo w erT1mean. Bysorting regionsinfeaturespacebasedupontheirmeanv alues,rulesbasedontheirrelativ e ordercanbecreated: 1.Largeregionsthatcon taintumorwilllik elycon tainasignican tn um berof pixelsthatareofhighestin tensit yinT1andPDspace. 2.Regionsthatdonotcon taintumorareunlik elytocon tainasignican tn um ber ofpixelsthatareofhighestin tensit yinT1andPDspaceandmostlik elyto con tainasignican tn um berofpixelsoflo w estin tensit yinT1andPDspace. 3.Thein tra-cranialregionwiththehighestmeanT1v alueanda\high"PD andT2v alue,isconsidered\FirstT umor,"againstwhic hallotherregionsare compared. 4.Themeansofregionsthatcon tainsimilartissuet ypeswillneigh boroneanother infeaturespace. 5.Otherregionsthatcon taintumorarelik elytofallwithin1to1.5standard deviations(dependingonregionsize)ofFirstT umorinT1andPDspace.

PAGE 140

125 (a) (b) (c) (d) (e) (f) Figure38.UsingPixelCoun tstoRemo v eNon-T umorousRegions.Giv enarened tumorsegmen tationafterStageThree(a)andremo v alofan yextra-cranial pixels,regionswithasignican tn um berofpixelshighestinT1space(b) orPDspace(c)arelik elytocon taintumor,((b)and(c)aremergedin (d)),whileregionswithpixelslo w estinT1space(e)areunlik elytocon tain signican ttumor.Groundtruthissho wnin(f).

PAGE 141

126 T able5.RegionLabelingRulesBasedonPixelPresence. An y Region a (with Size a )thatsatisesoneofthefollo wingrulesislabeledandremo v ed fromfurtherconsideration. Sizeof PixelsinIn tersection Action Region a with3Masks 5 An yBottomT1PixelsAND Remo v e Lessthan2T opT1Pixels Non-tumor 500 Morethan Size a 0 : 06T opT1Pixels LabelAs T umor 5 NoT opT1PixelsAND Remo v e >Size a 0 : 005BottomT1PixelsAND Non-tumor
PAGE 142

127 T able6.Remo vingNon-T umorRegionsBasedonStatisticalMeasuremen ts.Regionsthatsatisfyoneofthefollo wingrulesareremo v edfromthetumor mask. Region Largest isthelargestkno wntumorregion. (a)RulesBasedonStandardDeviation(SD)of\FirstT umor" RegionSize Remo v eifRegion'sMeanV aluesare: 10OR > 1SDa w a yinT1spaceOR Region Largest = 4 > 1SDa w a yinPDspace. 10AND > 1 : 5SDa w a yinT1spaceAND Region Largest = 4 > 1 : 5SDa w a yinPDspace. (b)LabelingRulesBasedonRegionStatistics 100 Region Skewness T 1 0 : 75AND Region Skewness PD 0 : 75AND Region Skewness T 2 0 : 75 tumorregionisv eriedviatheheuristicthatatumorregionwillnotonlyha v eav ery highT1meanv alue,butwillalsooccup ythehighesthalfofallregionsinsortedPD andT2meanspace. F orexample,iftherew ere10regionstotal,theregionbeing testedm ustbeoneofthe5highestmeanv aluesinbothPDandT2space.Ifthe candidateregionpasses,itisconrmedasFirstT umor.Otherwise,itisdiscarded andtheregionwiththenexthighestT1meanv alueisselectedfortestingasFirst T umor. OnceFirstT umorhasbeenconrmed,extractedkno wledgeindicatesthatregionsofasimilartissuet ypeneigh boroneanotherinfeaturespace. Althoughtumorousregionscanha v esignican tlydieren tmak eupsbet w eenslices,thisheuristic holdsforthepurposeofseparatingtumorfromnon-tumorregionswithinaslice. F urthermore,thestandarddeviationsinT1andPDspaceofakno wntumorregion w erefoundtobeausefulandexibledistancemeasure. T able6(a)liststhet w orulesthatusethestandarddeviationtoremo v enontumorregions,basedonthesizeoftheregionbeingtested.TheruleinT able6(b) serv esasatie-break erforsomeregionsthatw erenotlabeledbefore. Theterm

PAGE 143

128 Region Largest isusedtoindicatethelargestkno wntumorregion.Inmostcasesthere w asonlyasingletumorregion,sotheFirstT umorregionw asalsothelargestregion. Incaseswheretumorw asfragmen ted,ho w ev er,alargertumorousregionwillpro vide amorerobustmeanandstandarddeviationforthedistancemeasure. Therefore, theterm Region Largest w ouldbeassignedtothelargestregionwithinonestandard deviationinbothT1andPDspacetotheFirstT umorregion. AftertherulesinT able6areapplied,thetumorsegmen tationwillha v eanal thresholdapplied. 5.5 StageFiv e:FinalT1Thresholding A ttheendofStageF our,theregionswithnotumorha v ebeenremo v ed,but non-tumorpixelsma ystillbefoundinthoseregionsconsideredtocon taintumor. Whileenhancingtumorhaspropertiesineac hofthethreea v ailablefeaturesthatha v e beenusedaskno wledge,discussionswithanexpertradiologist[84]ha v eindicatedthat naltumorboundariesaredeterminedb ypixelin tensitiesintheT1-w eigh tedimage. Thresholdsw eredescribedinSection5.2inarelativ ely\coarse"mannerbecause theboundaryofenhancingtumorw as\obscured"b ypixelsbelongingtonon-tumor tissues.Withtheremo v alofmostofthesenon-tumortissuesinStagesTw othrough F our,ho w ev er,agreaterlev eloffocuscanbeplacedandamoreprecisethreshold canbeapplied. Kno wledgepresen tedinSection5.4.2indicatedthatspatialcomponen tscontainingtumorha v eameanT1v aluewithinappro ximatelyonestandarddeviationof themeanT1v alueofkno wntumor.Applyingthisheuristicatapixellev elsuggests thatausablethresholdcouldbesetatthemeanv alueofthesegmen tedtumorimage min usitsstandarddeviationinT1-w eigh tedspace. Whileeectiv einsomeslices, thesignican tdegreeofv ariabilit yoftumor,intermsofsize,o v erallsignalin tensit y

PAGE 144

129 inT1-w eigh tedspace,anddegreeofseparabilit yfromsurroundingnon-tumortissues prev en tsasinglethresholdform ulafrombeingused.Therefore,thethresholdismade adaptiv eb ygeneratingaseriesofcandidatev alues,basedonthe\standarddeviation" principlestatedabo v e,andc hoosingthethresholdthatbestappro ximatesthetumor boundaryb yexaminingthemak eupofthecurren ttumorsegmen tationimage. Thecandidatethresholdsaregeneratedb yrstexaminingho wthepixelscontainedwithinthetumorsegmen tationmaskproducedattheendofStageF our(hereaftercalledthe\T otalT umor"mask)w erepartitionedduringinitialF CMsegmentation.A\classmap"ofthetumorsegmen tationiscreatedloadingtheinitialF CM segmen tationimageandk eepingonlythosepixelscon tainedintheT otalT umormask. Thiscreatesanimagethatisspatiallyiden ticaltotheT otalT umormask,butalso con tainstheclasslabelsofeac hpixelfromtheF CMsegmen tation,pro vidingadditionalkno wledge,bothforgeneratingthecandidatethresholdsandselectingthemost appropriateone.Examplesaresho wninFigures39(a)and(e). Theterm Largest isassignedtothe Cluster i thathasthegreatestn um berof pixelscon tainedb ytheclassmapimage.Thepixelsbelongingto Cluster Largest are separatedfromtheT otalT umormask,creatinga\LargestCluster"mask(notto beconfusedwiththe\LargestT umor"spatialregiondenedinStageF our),while a\RemainingClusters"maskcon tainingallremainingpixelsfromtheT otalT umor mask.ExampleofLargestClustermasksaresho wninFigures39(b)and(f),while (c)and(g)sho w\RemainingClusters"masks. OncetheLargestClustermaskhasbeencreated,ho wtumorw aspartitioned duringF CMclusteringcanbedetermined.Enhancingtumoriskno wntoha v ethe highestT1v alueofallin tra-cranialtissues.Therefore,if Cluster Largest con tainsthe majorit yofthetumor,thentheLargestClustermaskshouldha v eameanT1v alue greaterthanthatoftheT otalT umormask.Otherwise,ifthetumorw assegmen ted in tom ultipleclassesduringF CMclustering,orifif Cluster Largest con tainsprimarily

PAGE 145

130 (a)Ra wData (b)Classmap Image (c)Pixelsin LargestCluster (d)Pixelsin RemainingClusters (e)PixelsRemo v ed AfterThresholding (f)Ra wData (g)Classmap Image (h)Pixelsin LargestCluster (i)Pixelsin RemainingClusters (j)PixelsRemo v ed AfterThresholding Figure39.FinalT1Thresholding. A\classmap"image,(b)and(g),isgenerated b ymappingtheF CMlabelson tothetumormask.Theclusterwiththe largestn um berofpixels,(c)and(h),isisolatedfromtherestofthemask, (d)and(i).TheLargestClusterimagein(c)con tainsthemostenhancing tumorsin(a),whilethemostenhancingtumorpixelsin(f)arefoundthe RemainingClustersmask(i).Basedonthisdistinction,statisticalrules areusedtosetanalthresholdinT1spaceallo wingadditionalnon-tumor pixelstoberemo v ed(e)and(j).

PAGE 146

131 non-tumorpixels,thentheLargestClustermaskma yha v eameanT1v aluelo w er thanthatoftheT otalT umormask. Thisconceptismeasuredb ycalculatingthemeanT1v alue, u T 1 ,asdened inSection5.4.2,fortheLargestClustermask( u T 1 Largest )andtheT otalT umormask ( u T 1 Total )andcreatingan\enhancemen tratio," Enhance T 1 L=T ,usingthefollo wingform ula: Enhance T 1 L=T = u T 1 Largest u T 1 Total (5.5) IftheLargestClusterimagehasagreatermeanT1v aluethantheT otalT umor image, Enhance T 1 L=T willbe 1 : 0.Otherwise,theLargestClusterimagehasalo w er meanT1v alue.Asimilarenhancemen tratio, Enhance T 1 L=R isgeneratedbet w eenthe LargestClustermask, u T 1 Largest ,andRemainingClustersmask, u T 1 Remaining : Enhance T 1 L=R = u T 1 Largest u T 1 Remaining (5.6) Theratiosareusedtoallo wtherulesinSections5.5.1and5.5.2togaugeho wdistinct theLargestClustermaskis(orisn't)fromtheothertumormasksandguidethe selectionofthenalthresholdv alue. Atotaloffourcandidatethresholdsaregenerated,threeofwhic harebasedon theT1meanandstandarddeviationofthethreetumormasks:T otalT umor,Largest Cluster,andRemainingClusters.Oneofthepossiblecandidatethresholds,described abo v e,canbedeterminedb ycalculatingthemeanT1v alueoftheT otalT umormask, u T 1 Total ,min usitsT1standarddeviation T 1 Total .Called\TMSD,"forT otal(T umor) Mean(Min us)StandardDeviation,itisdenedas: TMSD = u T 1 Total )Tj/T17 1 Tf50 0 TD( T 1 Total (5.7)

PAGE 147

132 T able7.CandidateThresholds.Basedonthestatisticalinformationofthetumor masks,aseriesofpossiblethresholdsinT1spacearegenerated.Thebest thresholdforaparticularsliceisthenselectedaccordingtorulesinthe kno wledgebase. Label MaskBasedOn ThresholdF orm ula TMSD T otalT umor u T 1 Total )Tj/T17 1 Tf50 0 TD( T 1 Total LMSD LargestCluster u T 1 Largest )Tj/T17 1 Tf50 0 TD( T 1 Largest RMSD RemainingClusters u T 1 Remaining )Tj/T17 1 Tf50 0 TD( T 1 Remaining WMT WhiteMatter T 1 WM ThresholdsfortheLargestClustermask(LMSD)andtheRemainingClustersmask (RMSD)aregeneratedsimilarlyusingtheirrespectiv eT1meansandstandarddeviations.ThesethresholdsaresummarizedinT able7. Thefourthcandidatethreshold,called\WMT"forWhiteMatterThreshold, istheT1cen troidv alueofthewhitemattercluster, T 1 WM .Ifwhitemattersplitting hasoccurred,theGroup2clusterwiththelo w estT2v aluecen troidisusedbecause itcanbeconsisten tlyiden tied,regardlessofthedegreeofpathologypresen t,and itscen troidhasthehighestT1v alueofallclustersofnormaltissues. Enhancing pathologyhasahigherT1v aluethanwhitematter. TheWMTthresholdisused toremo v easetofpixelsonthewhitematter/pathologyborder. Then um berof pixelsremo v edwillnormallybesmall(justthosethataresligh tlyhigherinT1signal in tensit ythattissuekno wntobewhitematter). TheWMTthresholdisadjustedsligh tlywhen Enhance T 1 L=T 1 : 0.Sincethe pluralit yofthetumor,con tainedin Cluster Largest ,sho wssignican tenhancemen t, asdenedb y Enhance T 1 L=T ,theWMTthresholdisusuallytoolo w.Therefore,itis scaledupb ytheenhancemen tratio,thefollo wingform ula: WMT =max ( T 1 WM ;Min T 1 Total ) Enhance T 1 L=T (5.8)

PAGE 148

133 where Min T 1 Total istheT1-w eigh tedv alueofthepixelintheT otalT umormaskwith thelo w estT1-w eigh tedin tensit y InsomesliceswheretheRemainingClustersimagehasahighstandarddeviation,thecalculatedRMSDthresholdcanbelessthan Min T 1 Total ,meaningtheRMSD thresholdwillha v enouse.T oprev en tthis,RMSDisresettothemaxim umofRMSD, Min T 1 Total ,andthescaledWMT.Inmostcases,RMSDisalreadythemaxim umv alue. Oncethefourthresholdsaregenerated,theyaresortedindescendingorder. Thev alue Threshold 1 ,forexample,willbegiv enthehighestofthefourcandidate thresholds,while Threshold 2 willbeassignedthesecondhighest,andsoon.Using criteriasuc hasthev alueof Enhance T 1 L=T andtheorderofthecandidatethresholdsin T1-w eigh tedspace,thecandidatethresholdthatbestappro ximatesthetumor/nontumorboundary(inT1-w eigh tedspace)willbeselected.Dieren tsetsofruleswill beactiv atedbasedonthev alueof Enhance T 1 L=T andw eredev elopedb yexamining thetrainingsetandobservingwhic hthresholdga v ethebestresultsunderdistinct conditions.Therulesarepresen tedinorderofapplication. 5.5.1 ThresholdingIf Enhance T 1 L=T 1 : 0 SliceswheretheLargestClustermaskhasahighermeanT1v alueindicatet w o possibilities.Inmostslices,thetumorenhancedev enlyandw asclusteredin toasingle classduringinitialF CMsegmen tation.Insomeslices,ho w ev er,dieren tareasofthe tumorenhancedindieren tamoun tsandw eresplitin tom ultipleclassesduringF CM segmen tation,withtheLargestClustermaskcon tainingthepixelswiththemost enhancemen t. Inthesecases,thenalthresholdv alueisdeterminedbasedona com binationofthev alueoftheenhancemen tratio, Enhance T 1 L=T ,andthe\spread" (orrange)ofthethresholdscalculatedb y:

PAGE 149

134 Threshold Spread = LMSD Threshold 4 (5.9) where Threshold 4 isthecandidatethresholdwiththelo w estT1v alue. SincetheLargestClustermaskhasahighermeanT1v aluethantheT otal T umormask,LMSDshouldalsobethehighestcandidatethreshold, Threshold 1 andissointhemajorit yofcases.IftheLargestClustermaskofaslicehasalarge enoughT1standarddeviation, T 1 Largest ,ho w ev er,theresultingLMSDthresholdma y notbethehighestcandidate.Intheseslices,theWMTthresholdw asconsisten tly foundtobethehighestthresholdinstead.Thisdistinctionisnotedandusedinthe decisionmakingprocess.Theconceptof\thresholdspread"isimportan tasan um ber ofrulesallo wamaxim um\spread"bet w eenthethresholdselectedand Threshold 4 Therstt w orulesaddresseasilyiden tiablecases.Normally ,allofthecandidatethresholdsaresignican tlylo w erthanthemeanT1v alueofeac hoftherespectiv ethreetumormasks.Ifaslice'sTMSDorWMTthresholdisgreaterthanthe meanT1v alueoftheRemainingClustersmask(themaskwiththelo w estmeanT1 v alue),thetumorneedsonlyminim althresholdingunlesstheslicesho wssignican t enhancemen t,inwhic hcasethenalthresholdissettoTMSD. IF u T 1 Remaining < max ( TMSD;WMT ) THENIF Threshold 1 = LMSD AND Enhance T 1 L=T 1 : 10 THENSetThresholdtoTMSD ELSESetThresholdto Threshold 4 Iftheenhancemen tratiobet w eentheLargestClusterandRemainingClusters masks, Enhance T 1 L=R ,isextremelysmall,thenthetumorenhancedev enly ,butw as splitin toseparateclassesb yF CMalonganotherfeature(PDorT2).Intheseslices, thenalthresholdissettoTMSD. IF Enhance T 1 L=R < 1 : 001 THENSetThresholdtoTMSD

PAGE 150

135 Thefollo wingruleaddressessliceswhereLMSDw asnotthehighestcandidate thresholdduetothehighT1standarddeviationoftheLargestClustermask.Inmost ofthesecases,thehighestthreshold,WMT,w asselectedunlessthelo w estt w ocandidatethresholds(usuallyTMSDandRMSD)hadv erysimilarv alues,indicatingev en enhancemen tthroughoutthetumormaskwhic hshouldreceiv eaminim althreshold. IF Threshold 1 6 = LSMD THENIF j Threshold 3 )Tj/T17 1 Tf48.9999 0 TD[(Threshold 4 j < 3 : 0 THENT ak eMaximum Threshold i SuchThat Threshold i Threshold 4 < 1 : 002 ELSESetThresholdtoWMT TheremainingrulesinthissectionassumethatLMSDisthehighestcandidate threshold, Threshold 1 .IfLMSDisthehighestcandidatethreshold,andthethreshold spreadisv erysmall,thenthetumorenhancedev enly ,butthenalthresholdisset toLMSD,asopposedtoTMSD,toallo wsligh tlymoreaggressiv eremo v aloffalse positiv epixels. IF Threshold Spread 1 : 04 THENSetThresholdtoLMSD Ifthethresholdspreadisv erylarge,thentheLargestClustermasknotonly hasasignican tlyhighermeanT1v alue,buthasastandarddeviationsuc hthatthe resultingLMSDthresholdissignican tlyhigherthantheothercandidatethresholds. T odeterminewhetherthetumorw asfragmen tedduringF CMclustering,theT1 standarddeviationoftheT otalT umorimageandenhancemen tratio Enhance T 1 L=R to determinethemaxim umallo w ableratiobet w eenthenalthreshold, Threshold i ,and thelo w estthreshold, Threshold 4 ,whic hisusuallythewhitematterthreshold,and th uskno wnnon-tumortissue.

PAGE 151

136 IF Threshold Spread 1 : 275 THENIF Enhance T 1 L=R > 1 : 25 THENIF T 1 Total > 180 THENT ak eMaximum Threshold i SuchThat Threshold i Threshold 4 < 1 : 10 ELSET ak eMaximum Threshold i SuchThat Threshold i Threshold 4 < 1 : 15 ELSET ak eMaximum Threshold i SuchThat Threshold i Threshold 4 < 1 : 275 Thenalruleco v erstheremainingthresholdspreadv alues.F orsliceswitha thresholdspreadabo v e1 : 1,iftheslice'senhancemen tratio, Enhance T 1 L=R ,isgreater thanitsthresholdspread,thenmostofthetumoriscon tainedintheLargestCluster maskandtheLMSDthresholdisused.Otherwise,thetumorenhancedunev enlyand minim althresholding,basedonwhitematter,isapplied. Ifthethresholdspreadislessthan1 : 1,theTMSDthresholdisusedifthere islittledierencebet w eenitandthelo w estcandidatethreshold. Otherwise,the relationshipbet w eenWMTandtheremainingthresholdsisconsidered. IF 1 : 04 1 : 1 THENIF Threshold SpreadRMSD AND WMT Threshold 4 < 1 : 01 THENSetThresholdtoWMT ELSESetThresholdtoLMSD

PAGE 152

137 5.5.2 ThresholdingIf Enhance T 1 L=T < 1 : 0 Sliceswhere Enhance T 1 L=T < 1 : 0indicatethateithertheLargestClustermask w ascomprisedprimarilyofnon-tumorpixels,orhadatumorthatenhancedunev enly andw as\fragmen ted"in tot w oormoreclassesduringF CMsegmen tation,withthe majorit yofthepixelsplacedin totheLargestClustermaskha vingthelo w estenhancemen t. Ifthelev elof\enhancemen t"bet w eentheLargestClusterandT otalT umor masksissignican tlylessthan1 : 0,thentheLargestClustermaskmostlik elycon tains onlynon-tumorpixels,whiletheRemainingClustersmaskhastumorwithstrong enhancemen t.Inthesecases,theT1standarddeviationoftheT otalT umormask willbehighandtheLMSDthresholdwillbeused.Otherwise,theTMSDthreshold isused. IF Enhance T 1 L=T < 0 : 90 THENIF T 1 Total > 180 THENSetThresholdtoLMSD ELSESetThresholdtoTMSD Asstatedearlier,aslicewithanenhancemen tratiolessthan1 : 0hasatumor wheremostoftheenhancingtumorpixelsarecon tainedintheRemainingClusters mask,nottheLargestClustermask. Asaresult,theRemainingClustersmask willinsteadcon tainthepixelswiththemostenhancemen tandtheRMSDthreshold willusuallybethehighestcandidatethreshold.IftheT1standarddeviationofthe RemainingClustersmask, T 1 Remaining ,islargeenoughthatRMSDisnotthehighest candidatethreshold,ho w ev er,thenthefollo wingruleresifthehighestcandidate thresholdisTMSDorWMT:

PAGE 153

138 IF Threshold 1 = TMSD THENSetThresholdtoTMSD ELSEIF Threshold 1 = WMT THENSetThresholdtoWMT Intheremainingcases,slicesarethresholdedprimarilybasedonthen um ber ofcandidatethresholdsabo v ea\Floor"v alue,whic hisnormallysettotheTMSD thresholdunlessTMSDhasav aluelessthanornear(inT1space)thelo w estT1v alue pixelintheT otalT umormask: Floor =max ( Min T 1 Total +5 ;TMSD ).Then um ber ofcandidatethresholdsgreaterthan Floor isdeterminedandaspecicruleisred dependingonthen um bercoun ted.Ingeneral,thisn um berincreasesastheT1standarddeviationincreasesintheT otalT umormask(whic hlo w ersTMSD),suggesting ahigherpresenceofnon-tumorpixels,ortheT1standarddeviationsintheLargest ClusterandRemainingClustersmasksdecreases(lea vingLMSDandRMSDathigher v alues),suggestinghigherhomogeneit yoftissueswithintherespectiv emasks. Ifnooronlyonecandidatethresholdisgreaterthan Floor ,selectionisdeterminedbasedonwhic hcandidatethresholdhasthemaxim umv alue.Asstatedearlier, themaxim umcandidatethresholdisnormallyRMSDsinceitisbasedonthemask con tainingmostoftheenhancingtumorpixels.IftheRMSDthresholdisthehighest threshold,itisselectedunlesstheslicehasasignican tT1standarddeviationinthe T otalT umormask,inwhic hcaseminim althresholdingisapplied. IF Threshold 1 = RMSD THENIF T 1 Total > 125 THENSetThresholdto Threshold 4 ELSESetThresholdto RMSD IftheLMSDthresholdisthehighestthreshold,ho w ev er,duetoanextremely highT1standarddeviationintheRemainingClustersmask,thentheLMSDthresholdisselectedunlesstheslicesho wshighT1standarddeviationsinallthreetumor masksandthewhitematterthreshold,WMT,issignican tlybelo wtheFloorv alue, promptingminimalthresholdingtobeapplied.

PAGE 154

139 IF Threshold 1 = LMSD THENIF Threshold 4 = WMT AND WMT=Floor< 0 : 90 AND min ( T 1 Largest ; T 1 Remaining ; T 1 Total ) > 100 THENSetThresholdto Threshold 4 ELSESetThresholdto Threshold 1 Ift w ocandidatethresholdsaregreaterthan Floor ,themaxim um ,thenthe thresholdissettotheFloorv alue(usuallyTMSD)unlessthesliceshasasignican t T1standarddeviationintheT otalT umormaskorhasalo wenhancemen tratio, suggestingarelativ elysmalltumorsurroundingb yalargeamoun tofnon-tumor pixels.Inthesecases,themaxim umthresholdisused. IF T 1 Total > 150 OR Enhance T 1 L=T < 0 : 95 THENSetThresholdto Threshold 1 ELSESetThresholdto Floor Finally ,ifthreecandidatethresholds(LMSD,RMSD,andWMT)w eregreater than Floor ,thentheTMSDthresholdw asthelo w estthresholdduetoasignican tT1 standarddeviationintheT otalT umormask.Therelationofwhitematter(WMT) andtheLargestClustermask(LMSD)determineswhic hcaseexistsandwhatthe appropriatethresholdis.IftheLMSDthresholdisgreaterthanwhitematter,then theLargestClustermaskcon tainstumorandhassucien tseparabilit ytoallo w maxim umthresholding.Otherwise,thethresholdisbasedontheT otalT umormask. IF WMT
PAGE 155

140 5.6 P ostProcessing:T umorV ericationinUpperSlices Thetumorsegmen tationrulessearc hforgadoliniumenhancingtumor.Some oftheupperslices,ho w ev er,whileproperlybeingclassiedasabnormalduetothe presenceofotherpathologicaltissuessuc hasedema,didnotcon tainenhancingtumor asdenedb yradiologist-labeled\groundtruth."SincetherulesinStages1through 5assumethepresenceofenhancingpathology ,thenaltumorimageproducedb y Stage5willcon tainonlyfalsepositiv epixels. T ov erifythepresenceofenhancingtumor,av ariationoftheenhancingpathologytestfromSection4.4.3isappliedtothenaltumorimage,usingthehistogram peakinT1,PD,andT2featurespaceasthresholds.Theresultan timagesarein tersectedandthen um berofpixelsremainingiscomparedwiththen um berbeforethe thresholdsw ereapplied.Iflessthan5%ofthepixelsremain,thesliceisreclassiedas \noenhancing-tumorabnormal."Otherwise,thetumorsegmen tationisconrmed.

PAGE 156

141 CHAPTER6 RESUL TS 6.1 T rainingandT estSets AssummarizedinT able8,638slicesfrom63v olumesand25subjectsw ere a v ailableforprocessing.Oftheseslices,194w erefrom\v olun teer"normals,27w ere fromgadoliniumenhancednormals,and417w erefrompatien tscans. Ofthe417 patien tslices,176w ere\lo w erslices"and385w ereconrmed,b yradiologistpathology reports,toha v eglioblastoma-m ultiform etumor,withtheremaining32slicesfound immedi atelyabo v eorbelo wthetumormass. F romthea v ailabledatasets,three trainingsetsw erecreated,assho wninT ables9and10fornormalandabnormal slicesrespectiv ely .Thersttrainingsetw ascreatedforthepurposeofqualitativ ely modelinganddetectingpathologyinthelo w erslices,aspresen tedinChapter4,and included38normal(23fromv olun teersand15fromgadoliniumenhancednormals) and27abnormalslices. Theremainingt w otrainingsetsw eredesignedfordev elopingthetumorsegmen tationsystemandw ereth uscomprisedofonlyabnormalslices. Thesecond trainingsetused46slicesindev elopingStagesOnethroughF ourofthetumorsegmen tationsysteminChapter5,while64slicesw ereusedforthenalT1threshold stepinStageFiv e.

PAGE 157

142 T able8.SummaryofMRDataAv ailable. T ype T rain/T est #Subjects #V olumes #Slices V olun teer T rain 4 5 23 (Nogadolinium, T est 9 21 171 nopathology .) T otal 13 26 194 GadoliniumNormals T rain 2 2 15 (Receiv edgadolinium, T est 2 2 12 nopathology .) T otal 4 4 27 P atien t T rain 5 13 70 (Receiv edgadolinium, T est 3 20 347 con tainedpathology .) T otal 8 33 417 T otal T rain 11 20 108 T est 14 43 530 T otal 25 63 638 T able11liststhetotaln um berofslices,v olumesandsubjects(v olun teer, gadoliniumenhancednormals,andpatien t)unseenb ythesystemduringkno wledge basedev elopmen tforthepathologydetectionandtumorsegmen tationstages.The section\UnseenA tAn yProcessingStage"inT able11totalsthen um berofcases completelyunseenb yallthreeprocessingstagesinthesystem,foreac hofthethree subjectt ypes. Then um berofcompletelyunseenpatien tslicesislargerthanthe n um berofpatien tslicesduringtumorsegmen tationbecausesomeslicesfrompatien t scansw erefoundtocon tainnopathology(lyingabo v eorbelo wthetumormass). Theseslicesw erenotprocessedb ythetumorsegmen tation,butw erestillconsidered completelyunseen.Becauseallpatien tscanshadsomepathologypresen t,then umberofcompletelyunseenpatien tv olumesandsubjectsw erecalculatedb ycoun ting then um berofeac hunseeninbothpathologydetectionandalltumorsegmen tation stages. AsT able10sho ws,patien ts3,6,and8w erecompletelyunseenbeforeprocessing.Anadditionalcolumn,\OneSlice,"isalsoincludedinT able11toindicatethe

PAGE 158

143 T able9.NormalMRSliceDistribution.P aren thesisindicatethen um berofslices fromthatv olumethatw ereusedastrainingforpathologydetection.V olun teerw asscannedusingan um berofprotocols,witheac hprotocolbeing listedseparately V=V olun teer;GN=Gadoliniumenhancednormal; Rn=RepeatScann #SlicesExtractedfromV olume V olun teer Baseline R1 R2 R3 R4 T otal V1 6 6 V2 6 6 V3 7 7 V4a 8 8 V4b 8 8 V4c 7 7 8(1) 22(1) V5 7 7 V6 6(6) 10 6 7(1) 8 37(7) V7 11 8(8) 9 28(8) V8 7 7 V9 7 7 V10 7(7) 7 14(7) V11 7 7 V12 6 6 V13 8 9 7 24 GN1 7(7) 7(7) GN2 8(8) 8(8) GN3 5 5 GN4 7 7 T otal 221(38)

PAGE 159

144 T able10.AbnormalMRSliceDistribution.P aren thesisindicatethen um berofslices fromthatv olumethatw ereusedastrainingfor(P athologyDetection; T umorSegmen tationStages1-4;FinalThresholding,Stage5). Rn= RepeatScann. #SlicesExtractedfromP atien tV olume Scan P1 P2 P3 P4 P5 P6 P7 P8 Base 10 13 12 16 9 15 12 7 (-/-/13) (1/-/-) (8/9/9) (1/7/7) R1 11 14 12 15 10 12 (-/-/1) (4/14/14) (1/-/1) (2/-/-) R2 11 15 15 8 13 (-/-/2) (9/15/15) (-/-/1) R3 10 8 15 (1/1/-) R4 12 7 15 (-/-/1) R5 16 R6 15 R7 14 R8 14 R9 11 R10 14 R11 18 R12 18 T otal 54 42 24 46 42 187 12 7 (-/-/1) (4/14/29) (11/15/16) (11/10/11) (1/7/7)

PAGE 160

145 T able11.UnseenMRSliceDistribution.Then um berofslices,v olumes,andsubjectsnotseenduringkno wledgebasedev elopmen tarelisted. \1Slice" indicatesthen um berofadditionalv olumesandsubjectsfromwhic ha singleslicew asusedduringkno wledgebasedev elopmen t.GN=Gadoliniumenhancednormal. Processing Subject Num ber Num ber Num ber \1Slice Stage T ype Slices V olumes Subjects V ol./Subj." P athology Normal 171 21 9 2/1 Detectionin GN 12 2 2 0/0 Lo w erSlices P atien t 155 25 4 4/1 T umorStages1-4 339 27 4 2/1 T umorStage5 P atien t 321 22 3 4/1 AllT umorStages 321 22 3 2/1 UnseenA t Normal 171 21 9 2/1 An yProcessing GN 12 2 2 0/0 Stage P atien t 344 20 3 2/1 Sum 527 43 14 4/2 n um berofadditionalv olumesandsubjectsthathadonlyasinglesliceusedforduring dev elopmen tofthekno wledgebase.F orexample,onlyasingleslicefromP atien t1 w asconsideredduringthenalthresholdingstage.Thelargertrainingsetallo w ed c haracteristicsthatw erecommonacrosspatien ts,asw ellasc haracteristicsthatcould sho wthemostv ariation,tobeextractedandusedasheuristics.Sincethetraining setw asselectedtobestco v erthedomainoftumorc haracteristics(size,etc.),new patien tsthatarein troducedshouldfallin toanareaalreadyco v eredb ythetraining set. 6.2 ResultsforP athologyDetection Duringpathologydetectioninthelo w erslices,ofthe221\normal"slicesprocessedfromeitherv olun teerorgadoliniumenhancednormalv olumes,ononlyt w o

PAGE 161

146 (a)Ra wImage (b)GTT umor (c)Ra wImage (d)GTT umor (e)Ra wImage (f)GTT umor Figure40.F ailuresonP athologyDetection.Images(a)and(c)sho wslicesthatcontainedtumor,butw erepassedb ythesystemascon tainingnopathology Theirrespectiv egroundtruth(GT)tumorimagesaresho wnin(b)and (d).Otherslices,suc has(e),hadpathologydetected,butw asconsidered nottocon tainenhancingtumorb ythepost-processingstepforupper slices.ThecorrespondingGTimageissho wnin(f).

PAGE 162

147 T able12.SystemF ailurestoDetectT umorinLo w erSlices.Thesliceslistedhere w eremisclassiedas\normal"b ythepathologydetectionsystemdescribedinChapter4. P atien t Scan Slice T umorSize 3 R1 13 13 6 R5 13 41 R6 11 8 12 13 R11 07 8 08 83 slicesdidafailureoccur.Bothoftheseslicesw erefromtestv olumeGN3andw ere theresultoftheheadbeingsucien tlyrotatedwithintheMRcoilthatthemask symmetrytestsinSection4.4.1failedandcausedtheocularm uscles(suppliedb y gadoliniuminfusedblood)tobedetectedasenhancingpathology .Thedegreeofrotationinheadw ouldnormallydisqualifythev olumefromprocessingsinceoneofthe system'srequiremen tsisthatthesubjectbelooking\straigh tup"withinthecoil, butw asincludedduetothelimitedn um berofgadoliniumenhancednormalsa v ailable.Sincethev olumew asatestcase,ho w ev er,nokno wledgew asextractedfrom thev olumeanditcanberemo v edwithoutaectingthesystem. Thesystemfailedtoproperlydetectenhancingpathologyinsixof176abnormal lo w erslices,listedinT able12.F ouroftheseslicescon tainedv erysmalltumors(13 pixelsorless)withlittleornosurroundingpathology(edemaand/ornecrosis)to distortthequalitativ emodels,makingthemdiculttolocate,ev envisually The remainingt w oslices,ho w ev er,hadlargertumorsizes.TheslicefromRepeatScan5, sho wninFigure40(a),w asfragmen tedandnosingleregionw assucien tlylargeto bedetected,buttheslicefromRepeatScan11,sho wninFigure40(c),w ascompact andalthoughitdidnotcon tainm uc hsurroundingpathology ,w aslargeenoughthat itshouldha v ebeendetectedb ytherulesinSection4.4.3.2.Examinationrev ealed

PAGE 163

148 T able13.F ailuresintheP ost-ProcessingStageforUpperSlices.Thecolumn\F alse No-T umor"columnsho wsthen um berofgroundtruthtumorpixelslost inslicesthatw eremisclassiedb ythesystemasha vingonlynon-tumor pathology The\F alseT umor"columnlistthen um beroffalsepositiv e tumorpixelsinslicesmisclassiedb ythesystemascon tainingtumorwhen nogroundtruthtumorw aspresen t.Thev alueinthecolumnindicates then um beroffalsepositiv es.Rn=RepeatScann. P atien t Scan Slice F alse F alse V olume Num ber T umor No-T umor #Pixels #Pixels 2 R1 24 165 R2 24 172 6 R1 26 72 R4 23 33 R6 23 94 24 77 R12 23 26 24 4 8 Base 22 289 thattheT1-w eigh tedthresholdgeneratedinSection4.4.3.1w astoohighandhad remo v edmostofthetumorfromconsideration. Theslicew asneigh bored(both abo v eandbelo w)b yslicesthathadpathologyproperlydetected,soarulecouldbe constructedtoconsidersuc hcases. Thesystemdetectedpathologyintheremaining170lo w erslices,detectingtumorassmallassixpixels,butthesixfailuressuggestthatsomeadditionalkno wledge couldstillbeusedtobettersetthethresholdsinSection4.4.3.1forsmalltumorswith littleornosurroundingpathology .Thesystemalsomisclassiedt w oslicesasfalsely ha vingenhancingtumor.Examinationoftheserev ealedthatwhilethesesliceshad someenhancemen t,theyw erenotconsideredb ytheradiologisttocon taintumor. Somefailuresalsooccurredinthepost-processingstagetoseparateenhancing tumorfromnon-tumorpathologyinupperslices,whic harelistedinT able13.One

PAGE 164

149 slice,fromP atien t8,BaselineScan,w asspatiallydisjoin tfromthetumormass(in fact,theslicesimmedi atelyabo v eandbelo witw ereproperlydetectedb ythepostprocessingstepasha vingnoenhancingtumor),soanadditionalruleforconstructing thetumorv olumew ouldremo v eitfromfurtherconsideration.Theremainingslices w erefoundimmedi atelyabo v eorbelo wthetumormass.\F alseNo-T umor"slices, misclassiedasha vingonlynon-tumorpathology ,w ereduetothelev elsofenhancemen tintheseslices,plusthetumor'sfragmen tationandsize(oneslicehadatumor sizeof4pixels).Anexampleofoneoftheseslicesissho wninFigure40(e),along withtheradiologistlabeledgroundtruth.\F alseT umor"sliceshadareaswithsome lev elofenhancemen t,buttheradiologistdidnotconsiderthemtocon taintumor. 6.3 ResultsforT umorSegmen tation 6.3.1 Kno wledge-Basedvs.GroundT ruth T oev aluatetheperformanceofthetumorsegmen tationsystem,alltumorsegmen tationsgeneratedb ythekno wledge-basedsystemw erecomparedwith\groundtruth"tumorsegmen tationsthatw erecreatedb yradiologistlabeling[125]. Error w asfoundbet w eenthet w osegmen tations,bothfalsepositiv es(wherethesystem indicatedtumorouspixelswheregroundtruthdidnot)andfalsenegativ es(where groundtruthindicatedtumorouspixelsthatthesystemdidnot).T ocompareho w w ell(onapixellev el)thekno wledge-basedsegmen tationcorrespondedwithground truth,t w omeasuresw ereused.Therst,\percen tmatc h,"issimplythen um berof truepositiv esdividedb ythetotaltumorsize.Thesecond,iscalleda\correspondence ratio,"andw ascreatedtoaccoun tforthepresenceoffalsepositiv es.Itisdenedas: CorrespondenceRatio= T rueP os. )Tj/T20 1 Tf50 0 TD((0 : 5 F alseP os. ) Num berPixelsinGroundT ruthT umor (6.1)

PAGE 165

150 (a)Ra wImage (b)KBBefore (c)KBFinal (d)GTT umor Figure41.AF ailureofT umorSegmen tation.Image(b)sho wsatumorsegmen tation afterStageTw o,whenextra-cranialtissuesareiden tied. Thesystem normallyremo v estheocularm uscles,butfailstodosointhiscasedueto thesizedieren tialoftheregions.Asaresult,analtumorsegmen tation isproducedin(c),whic hdisagreeswithgroundtruth(d). Theratiowillha v ethemaxim umv alueof1when T rueP ositive = NumberofPixels inGroundT ruthT umo r and F alseP ositive = F alseNegative =0.Thepoorestpossible segmen tationwillha v earatioof: )Tj/T20 1 Tf44 -33.0001 TD(0 : 5 (#PixelsMRImag e-#PixelsinGroundT ruthT umor) #PixelsinGroundT ruthT umor (6.2) F orcomparingonaperv olumebasis,thea v eragev alueforP ercen tMatc hw as generatedusing: Av erage%Matc h= P slicesinset i =1 (%matc h) i (n um bergroundtruthpixels) i P slicesinset i =1 (n um bergroundtruthpixels) i (6.3) Thea v eragev aluefortheCorrespondenceRatioissimilarlygenerated. T able14liststheresultsofthekno wledge-basedsystemonaper-v olumebasis. Thekno wledge-basedsystemperformsw ello v erallwhere264ofthe385sliceskno wn

PAGE 166

151 T able14.Kno wledge-BasedT umorvs.RadiologistLabeledT umorP erV olume. P atien t Scan T rue F alse F alse T umor P ercen t Corr. P ositiv e P ositiv e Negativ e Size Matc h Ratio 1 Base 6958 1012 255 7213 0.96 0.89 R1 6940 1953 300 7240 0.96 0.82 R2 7215 1838 255 7470 0.97 0.83 R3 6221 1318 174 6395 0.97 0.87 R4 6120 1988 234 6560 0.96 0.85 2 Base 11277 4937 953 12230 0.92 0.72 R1 12688 2935 1756 14609 0.87 0.77 R2 18187 4068 2737 20924 0.87 0.77 3 Base 9926 928 966 10892 0.91 0.87 R1 5737 1650 234 5971 0.96 0.82 4 Base 9154 2545 1265 10454 0.88 0.75 R1 8672 1335 2163 10835 0.80 0.74 R2 13824 2300 1964 15788 0.88 0.80 5 Base 9675 1064 503 10178 0.96 0.93 R1 4537 1398 120 4657 0.97 0.82 R2 4993 1309 623 5616 0.89 0.77 R3 8524 1088 691 9215 0.93 0.88 R4 3198 1759 346 3544 0.90 0.65 6 Base 5524 5742 305 5829 0.95 0.46 R1 2634 6471 50 2684 0.98 -0.22 R2 3821 6000 533 4354 0.88 0.19 R3 6143 6091 367 6510 0.94 0.48 R4 7768 4676 920 8688 0.89 0.62 R5 2432 3075 647 3079 0.79 0.29 R6 3940 2522 444 4384 0.90 0.61 R7 3523 2124 524 4047 0.87 0.61 R8 3180 1669 755 3935 0.81 0.60 R9 2652 2280 549 3201 0.83 0.47 R10 3116 1746 779 3895 0.80 0.58 R11 5258 2988 620 5878 0.91 0.63 R12 6292 2982 1129 7421 0.83 0.52 7 Base 842 358 176 1018 0.83 0.65 8 Base 4309 530 212 4521 0.95 0.89

PAGE 167

152 tocon taingroundtruthtumorhadaP ercen tMatc hratingof90%orhigherand resultedin18of33tumorv olumesha vingaP ercen tMatc hof90%,with10at 95%orbetter.T umorpixelscouldbelostduringthein tra-cranialextractionstage inslicewherethetumoroccupiedtheperipheryofthebrain,suc hasinP atien t 4RepeatScan2. Infouruppermosttestslices(allfromP atien t1),partofthe tumorhadgro wnbey ondthein tra-cranialregionin toanareanormallyoccupiedb y surroundingmeningialmem branes,whic hha v eanincreasedpercen tagepresencein theuppermostslices.Thetumor'slocationwithinthesemem branes,com binedwith thereducedbrainsizecomplicatedextraction.T umorlossalsooccurredduetomore subtlegadoliniumenhancemen t(stilldetectedb ytheradiologist,butnotclearenough infeaturespace)[37],whenthetumorw asfragmen tedandenhancedv eryunev enly andthesystemfocussedontheareaswiththegreatestenhancemen t,orcaseswhere tissuenecrosisprev en tedcirculationoftheenhancingagen t,buttheradiologistw as conserv ativ einthediagnosis.Lastly ,ifextra-cranialpixelsin troducedduringStage Onew erenotsuccessfullyremo v ed,focuscouldbemisplaceduponthemb ythe systemandtheirhighT1-w eigh tedin tensit y(duetothehighlev elsofgadolinium infusedblood)w ouldcausethetumortobemissed.Anexampleofthisissho wnin Figure41.Inthisslice,fromP atien t6,RepeatScan5,theocularm usclesintheleft hemispherew ereconnectedtomeningialtissuessurroundingthebrain.Duetothe sizedieren tial,theruleforremo vingoculartissues,whic hrequiresthatoculartissues correspondingacrossthebrainhemispheresha v eappro ximatelythesamesize,didnot remo v ethisregionanditw asconsideredtumorduetoitsT1-w eigh tedin tensit y .This sliceistheonlyfailureoftherulesinStageF our,Section5.4.1.1,buttheproblem shouldbeaddressed. Ov erall,thekno wledge-basedapproac htendedtoo v erestimatethetumorv olume.Onlyonev olumeinT able14sho wsnoticeableunderestimationb ythekno wledgebasedsystem(P atien t4,RepeatScan1),whic hcanbetracedtothreeslices.Tw o

PAGE 168

153 oftheseslices,hadpoorin tra-cranialmaskpreparationinStageOne,whileonthe remainingslice,apoornalT1thresholdw asc hoseninStageFiv e. Inallother cases,thekno wledge-basedsystemconsisten tlysho wsasignican tamoun tof\false positiv es."Sincethesystemsegmen tstumorb yremo vingonlypixelspro v ennotto betumor,lea vingan ythingthatremainsbeinglabeledastumor,thenahigherlev el offalsepositiv esisnotinconsisten twiththeparadigm. Examplesofkno wledge-basedsegmen tationv ersusground-trutharesho wnin Figure42tovisuallysho wthekno wledge-basedsystemtumorcorrespondenceto radiologist-labeledtumor.Figures42(a)through(c)sho waw orstcasesegmen tation, while(d)through(f)and(g)through(i)sho wana v erageandbestcasesegmen tation respectiv ely .Allthreeexamplesarefromslicesnotinthetrainingset. 6.3.2 Kno wledge-Basedvs.SupervisedMethods Oneoftheadv an tagesofthiskno wledge-basedapproac histhath umanbased trainingregionsofin terest(R OI's)perslice,curren tlyrequiredforallsupervised tec hniques[115],arenolongernecessaryafterruleacquisition.Y et,resultscanbeas good,ifnotbetter,thanthoseobtainedfromsupervisedmethods,withouttheneedto fortime-consumingR OIselection,whic hmak esuc hmethodsimpracticalforclinical useanddonotguaran teesatisfactoryperformance.T odemonstratethis,T able15 sho wsho ww ellthek-nearestneigh bors(kNN)algorithm(k=7)[28]performedonthe sameslicesprocessedb ythekno wledge-basedsystem.ThesupervisedkNNalgorithm ndsthek=7labeledR OIpixelsclosesttoatestpixelandclassiesthetestpixel in tothemajorit yclassoftheassociatedR OI's.ThekNNalgorithmhasbeensho wn, inotherstudies,tobelesssensitiv etothec hoiceofR OI'sb ydieren tobserv ersthan seed-gro wing[115,114].

PAGE 169

154 (a)Ra wImage (b)KBT umor (c)GTT umor (d)Ra wImage (e)KBT umor (f)GTT umor (g)Ra wImage (h)KBT umor (i)GTT umor Figure42.Kno wledge-BasedT umorSegmen tationvs. GroundT ruth. W orstcase (a-c),a v eragecase(d-f),andbestcase(g-i).

PAGE 170

155 T able15.kNNT umorvs.RadiologistLabeledT umorP erV olume.(k=7). P atien t Scan T rue F alse F alse P ercen t Corr. P ositiv e P ositiv e Negativ e Matc h Ratio 1 Base 6430 3592 782 0.89 0.64 R1 6548 5410 781 0.89 0.52 R2 6544 5032 925 0.88 0.54 R3 5643 5227 751 0.88 0.47 R4 5457 6657 1101 0.83 0.32 2 Base 8941 5473 3289 0.74 0.51 R1 7647 8331 6963 0.52 0.29 R2 10347 10339 10577 0.49 0.25 3 Base 8012 6077 2923 0.73 0.46 R1 4287 4150 1679 0.72 0.37 4 Base 6594 14968 3858 0.63 -0.13 R1 6083 15203 4751 0.56 -0.14 R2 10940 14067 4845 0.69 0.15 5 Base 8128 5914 2049 0.79 0.51 R1 3145 7134 1512 0.68 -0.09 R2 4421 14182 1204 0.79 -0.47 R3 7726 11543 1960 0.84 0.21 R4 1823 17315 1712 0.51 -1.90 6 Base 3605 9496 2211 0.62 -0.19 R1 2414 6936 340 0.88 -0.38 R2 4183 4898 171 0.96 0.40 R3 6500 386 10 1.00 0.97 R4 7565 7963 1123 0.87 0.41 R5 2501 3024 578 0.81 0.32 R6 3421 2753 963 0.78 0.47 R7 3318 4460 729 0.82 0.27 R8 3152 2125 783 0.80 0.53 R9 2729 1532 472 0.85 0.61 R10 3417 4847 478 0.88 0.26 R11 3930 5324 1922 0.67 0.22 R12 3811 4561 3610 0.51 0.21 7 Base 874 1490 144 0.86 0.13 8 Base 2412 2109 1733 0.53 0.30

PAGE 171

156 Itm ustbenotedthatmostofthekNNresultsincludeextra-cranialpixels inthetumorclassbecausekNNisappliedtothewholeimage(thereisnoprior in tra-cranialregionextraction).SomeofthesekNNv olumeshadthetumorman uallyextractedin[116,126,117](th usremo vingan yextra-cranialpixels,whic hthe kno wledge-basedsystemdoesautomatically)andaresho wninT able16. F urthermore,whilethekno wledge-basedsystemw asbuiltfromasubsetofthea v ailable slices,thekNNn um bersw erethemeanresultso v erm ultipletrialsofR OIselection oneac hslice.ThismeansthatallkNNsegmen tationsw eretrainingslicesandintroducesthequestionofin terandin tra-observ erv ariabilit y .Thekno wledge-based system,ho w ev er,processedan um berofslices,v olumes,andsubjectsinunsupervised modewithastaticrulesetallo wingforcompleterepeatabilit ywithnoin traand in ter-observ erv ariabilit y T able16comparesthetotaltumorv olumeofgroundtruth,thekno wledgebasedmethod,andkNNonpatien tswhereresultsw erea v ailableforbothmethods. ThekNNv olumessho wnareameano v eroneormoretrialso v eroneormoreobserv ers andincludethetotal(in terandin tra-observ er)standarddeviation.F ortheman ually extractedtumorkNNv olumes,the\best"v olumew asselectedforman ualextraction. Acomparisonisalsomadeagainstthesemi-supervisedF CM(ssF CM)algorithm,whic hw asinitializedwiththesameR OI'susedtoinitializekNN[116,126,117]. Theresultan tssF CMsegmen tationw asthenusedtoinitializeISG,acommerc iall y a v ailableseed-gro wingtool(ISGT ec hnologies,T oron to,Canada)forsupervisedev aluationoftumorv olumes.TheISGprocessingalsoremo v edan yextra-cranialtissues foundinthessF CMsegmen tation.Resultsa v ailableforthev olumesprocessedb y ssF CMandISGaresho wninT able17.Lik ethekNNv olumes,theresultsreported inT able17areameano v erthesetoftrialsperformedforthatv olumeandth usha v e astandarddeviation,alsolisted.

PAGE 172

157 T able16.Kno wledge-BasedT umorvs.kNN.(P at.=P atien t;GT=Groundtruth v olume;KB=Kno wledge-based;kNNSD=kNNstandarddeviation; Man ualkNN=kNNv olumeafterman ualtumorextraction;#T rial= Num beroftrials;#Obs.=Num berofobserv ers;N/A=Notapplicable/a v ailable.) P at. Scan GT KB kNN kNN Man ual # # V ol. V ol. V ol. SD kNN T rials Obs. 1 Base 7213 7910 10022 732 6334 5 2 R1 7240 8893 11958 2236 6794 5 2 R2 7470 9281 11576 1615 6616 5 2 R3 6395 7539 10870 4395 5901 5 2 R4 6560 8314 12115 891 5690 5 2 2 Base 12230 16214 14414 1257 N/A 5 2 R1 14609 15623 15979 2483 N/A 5 2 R2 20924 22255 20687 2622 N/A 5 2 3 Base 10892 10854 14090 2045 N/A 5 3 R1 5971 7387 8438 2361 N/A 5 3 4 Base 10454 11827 21564 3391 N/A 4 2 R1 10835 10007 21287 5944 N/A 4 2 R2 15788 16124 25008 3266 N/A 4 2 5 Base 10178 10739 14044 1901 7938 4 2 R1 4657 5935 10279 3242 2834 4 2 R2 5616 6302 18603 1084 3952 4 2 R3 9215 9612 18210 2685 6729 4 2 R4 3544 4957 19138 3789 3035 4 2 6 Base 5829 11266 13101 N/A N/A 1 1 R1 2684 9105 9350 N/A N/A 1 1 R2 4354 9821 9081 N/A N/A 1 1 R3 6510 12234 6886 N/A N/A 1 1 R4 8688 12444 15528 N/A N/A 1 1 R5 3079 5507 5525 N/A N/A 1 1 R6 4384 6462 6174 N/A N/A 1 1 R7 4047 5647 7778 N/A N/A 1 1 R8 3935 4849 5277 N/A N/A 1 1 R9 3201 4932 4261 N/A N/A 1 1 R10 3895 5081 8264 N/A N/A 1 1 R11 5878 8246 9254 N/A N/A 1 1 R12 7421 9274 8372 N/A N/A 1 1 7 Base 1018 1200 2364 N/A N/A 1 1 8 Base 4521 4839 4145 N/A N/A 1 1

PAGE 173

158 T able17.Kno wledge-BasedT umorvs. ssF CMandISG.ResultsforP atien t6,7, and8forssF CMandISGw ereuna v ailable. (P at. =P atien t;GT= GroundT ruthV olume;KB=Kno wledge-based;kNNSD=kNNStandardDeviation;#T rial=Num berofT rials;#Obs.=Num berofkNN Observ ers;N/A=NotAv ailable.) P at. Scan GT KB ssF CM ssF CM ISG ISG # # V ol. V ol. V ol. SD V ol. SD T rials Obs. 1 Base 7213 7910 8015 540 6067 303 5 2 R1 7240 8893 7757 1435 5956 177 5 2 R2 7470 9281 7362 229 6087 198 5 2 R3 6395 7539 7185 639 5361 121 5 2 R4 6560 8314 7332 1378 5172 210 5 2 2 Base 12230 16214 12457 716 10027 599 5 2 R1 14609 15623 10916 1371 7120 1084 5 2 R2 20924 22255 13498 864 10120 608 5 2 3 Base 10892 10854 8018 632 N/A N/A 5 3 R1 5971 7387 5337 151 N/A N/A 5 3 4 Base 10454 11827 8563 399 6258 146 4 2 R1 10835 10007 8901 390 6040 179 4 2 R2 15788 16124 14080 775 10819 1884 4 2 5 Base 10178 10739 10334 633 7562 645 4 2 R1 4657 5935 3813 554 3142 745 4 2 R2 5616 6302 4667 404 3483 229 4 2 R3 9215 9612 7852 185 7300 416 4 2 R4 3544 4957 3110 576 2311 306 4 2

PAGE 174

159 ComparedagainstkNN(beforeman ualextraction),thekno wledge-basedsystemisclosertothegroundtruthv olumein26of33v olumes,with10ofthemb y morethanthestandarddeviationofthekNNv olume(whenm ultipletrialsw ereperformed).Asignican tn um berofthepixelsintheoriginalkNNsegmen tationsare duetoextra-cranialtissues,assho wnwhencomparingthekNNv olumeswithand withoutman ualtumorextraction,assho wninT able16.Comparingtheirrespectiv e CorrespondenceRatios(onaper-v olumebasis),thekno wledge-basedsystemperformedbetterin29ofthe33patien tv olumes(88%),whilekNNperformedbetterin 4(12%).Moreimportan tly ,asT able15sho ws,thekNNmethodalsohasasignifican tlyhighern um beroffalsenegativ esin26of33v olumes(79%).F orkNNafter man ualtumorextraction,onlytotaltumorv olumes(ofthebestsegmen tations)w ere a v ailable,soasimilarcomparisoncouldnotbemade. ThessF CMapproac hperformsbetterthanthekno wledge-basedmethodin 10outof18v olumes,butonly3ofthesecasesw ereb ymorethanthestandard deviationofthessF CMv olume.Inthe8caseswherethekno wledge-basedmethod ga v ebetterresults,ho w ev er,7ofthemw erebetterthanssF CMb ymorethanthe standarddeviation.Thekno wledge-basedmethodperformsbetteragainstISG,in11 outof16cases,withall11b ymorethanthestandarddeviationofISG.F urthermore, ssF CMunderestimatedtotaltumorv olumein12instances,whileISGunderestimated tumorv olumeinall20a v ailablev olumes,whic hisnothelpfulforan yusein v olving treatmen t. 6.3.3 Ev aluationOv erRepeatScans Examiningtumorgro wth/shrink ageo v errepeatscans,thekno wledge-based methodfailedtoproperlytrac kthreeof25transitions,P atien t2(BaselinetoRepeat Scan1),P atien t4(BaselinetoRepeatScan1),andP atien t6(RepeatScans8to9).

PAGE 175

160 ThekNNmethod,withoutman ualtumorextraction,failedoneigh tof25transitions, whiletheman uallyextractedkNNv olumesfailedint w oof10transitions.ThessF CM methodfailedonthreeofthirteentransitions,whileISGfailedonfouroutoft w elv e. SincethekNN,ssF CM,andISGv olumesarebasedonm ultipletrials,itisdicultto assignaspeciccause,althoughtheimportanceofsupervisedremo v alofextra-cranial tissues,aprocesshandledautomaticallyb ythekno wledge-basedsystem,shouldbe noted.Also,asapercen tage,thekno wledge-basedsystemhadalo w errateoffailure thantheotherfourmethod. Theexplanationforthekno wledge-basedsystem'sfailuretocorrectlypredict tumorgro wthinP atien t2,fromtheBaselinescantoRepeatScan1hasat w o-fold origin.Accordingtopathologyreports,theBaselinescanhadasignican tamoun t ofuid,possiblyhemorrhage,whic hlefttobrigh tenedregionssurroundingthetumor inthePDscanandmadetheborderbet w eennon-tumorandtumorpixelsun usually diuse.Thisdistortedthehistogramonwhic htheinitialtumorsegmen tationw as based,resultinginsignican to v erestimationoftumorv olumeontheBaselinescan. InRepeatScan1,ho w ev er,notonlyhadtheuiddisappeared,butpathologyreports notedthesligh tdecreaseingadoliniumenhancemen t.Th us,theinitialo v erestimation follo w edb ythedecreasedgadoliniumenhancemen tcausedthetrendtoappeartobe tumorshrink ageinsteadofgro wth.AlthoughthekNNmethod(beforeman ualtumor extraction)follo w edthetrend,thebreakdo wnofho wman yfalsepositiv esw ereextracranialpixelsandho wman yw erewithinthetumorbedw asuna v ailable.Lookingat P atien t2inT able15,ho w ev er,then um berofT rueP ositiv esdecreases,suggesting thatthekNNmethodma yha v efailedtofollo wthetrendalso.F urthermore,both thessF CMandISGmethodsalsomissedthetransition. Infact,areviewofthe pathologyreportssho w edthatradiologistestimationsofthetumorv olumehadtobe revised.

PAGE 176

161 5000 6000 7000 8000 9000 10000 11000 12000 13000 0 1 2 3 4 Tumor Volume (voxels)Scanning Session Knowledge Based Versus Hand LabelingKB GT kNN (before) kNN (after) ssFCM ISG P atien t1 6000 8000 10000 12000 14000 16000 18000 20000 22000 24000 0 1 2 Tumor Volume (voxels)Scanning Session Knowledge Based Versus Hand LabelingKB GT kNN (before) ssFCM ISG P atien t2 Figure43.T rac king T umor Response Ov er Repeat Scans, P atien ts 1 and 2. KB=Kno wledge-BasedSystem; kNN=k-NearestNeigh bors; (before)=Beforeman ualtumorextractiontoremo v eextra-cranialpixels;(after)=Afterman ualtumorextraction.GT=GroundT ruth.

PAGE 177

162 5000 6000 7000 8000 9000 10000 11000 12000 13000 14000 15000 0 1 Tumor Volume (voxels)Scanning Session Knowledge Based Versus Hand Labeling KB GT kNN (before) ssFCM P atien t3 6000 8000 10000 12000 14000 16000 18000 20000 22000 24000 26000 0 1 2 Tumor Volume (voxels)Scanning Session Knowledge Based Versus Hand Labeling KB GT kNN (before) ssFCM ISG P atien t4 Figure44.T rac king T umor Response Ov er Repeat Scans, P atien ts 3 and 4. KB=Kno wledge-based system; kNN=k-nearest neigh bors; (before)=Beforeman ualtumorextractiontoremo v eextra-cranialpixels;(after)=Afterman ualtumorextraction.GT=GroundT ruth.

PAGE 178

163 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000 0 1 2 3 4 Tumor Volume (voxels)Scanning Session Knowledge Based Versus Hand Labeling KB GT kNN (before) kNN (after) ssFCM ISG P atien t5 2000 4000 6000 8000 10000 12000 14000 16000 0 1 2 3 4 5 6 7 8 9 10 11 12 Tumor Volume (voxels)Scanning Session Knowledge Based Versus Hand LabelingKB kNN(before) GT P atien t6 Figure45.T rac king T umor Response Ov er Repeat Scans, P atien ts 5 and 6. KB=Kno wledge-based system; kNN=k-nearest neigh bors; (before)=Beforeman ualtumorextractiontoremo v eextra-cranialpixels;(after)=Afterman ualtumorextraction.GT=GroundT ruth.

PAGE 179

164 (a)Ra wImage (b)KBT umor (c)GTT umor (d)F alseP ositiv es Figure46.T umorOv erestimation. Thekno wledge-basedsystem(segmen tation sho wnin(b))generallyo v erestimatedtumorincomparisontoground truth(c). Image(d)sho wsareasoffalsepositiv esingra y ,whic hsho w someenhancemen tintheT1-w eigh tedra wimage(theleftmostimagein (a)).Whilenotenhancingtothedegreeofpixelscon tainedb ytheground truthimage,theradiologistw ouldnotruleoutthepresenceofenhancing tumor,insteadsuggestingthelik elihoodofmicro-inltration. Thekno wledge-basedsystem'sfailureinP atien t4fromtheBaselinescanto RepeatScan1alsohast w osources. Thefalsenegativ esinRepeatScan1w ere discussedabo v e,andtheotherfactorcanbeprimarilyattributedtoasingleslice intheBaselinescanwhereapieceofmeningialtissueconnectsanextremelylarge extra-cranialregion(appro ximately900pixels)tothein tra-cranialmask,prev en ting itsremo v alinStageF our.Remo vingthisextra-cranialregionw ouldreducethetumor v olumeestimationtomak ethetransitionstatisticallyinsignican t. ThefailureonP atien t6w asduetoo v erestimationinRepeatScan9,found primarilyint w oslicesin tersectingthev en tricles,andth usindicatedasligh tv olume increase,wherethegroundtruthimageindicatedshrink age.Theseslices,ho w ev er, alongwithotherslicesfromP atien t6withsimilaro v erestimation,asdenedb y

PAGE 180

165 groundtruth,w erepresen tedtoaradiologistforev aluation.Inallcases,theradiologistconsideredthekno wledge-basedsystemtoha v ecapturedareaswhic hcouldha v e signican tmicro-inltration[85].Infact,whilenotincludedinthegroundtruthfor v olumedetermination,suc hareasw ouldlik elybeincludedfortreatmen tpurposes. Anexampleofoneoftheseslicesissho wninFigure46. Theprocessofcreatinggroundtruthimagesisv eryimprecise[85]andha v eappro ximatelya5%in ter-observ erv ariabilit yintumorv olume[125].Allbraintumors ha v emicro-inltrationbey ondthebordersdenedwithgadoliniumenhancemen t. Thisisespeciallytrueinglioblastoma-m ultiform es,whic harethemostaggressiv e gradeofprimarygliomabraintumors,andnoonecantellthe exact tumorborders, ev enwithin v asiv ehistopathologicalmethods[23,37,86],whic hw ereuna v ailable. Groundtruthimagesmarktheareasoftumorexhibitingthemostangiogenesis(formationofbloodv essels,resultinginthegreatestgadoliniumconcen tration)andrepresen tthosepixelswhic hare\statisticallymostlik ely"tocon taintumor[85,86]. Suc hpixelsw ouldha v ethehighestlev elofagreemen tagreemen tbet w eenradiologists,buttheydonotguaran teethatalltumorhasbeeniden tied[85].Therefore, thekno wledge-basedsystemma yoftencapturetumorboundariesthatextendin to areassho winglo w erdegreesofangiogenesis(whic hw ouldstillbetreatedduringtherap y)[86]. Figures43through45graphicallycomparethetotaltumorv olumesforthe repeatscansusingkno wledge-based,kNN(beforeandafterman ualtumorextraction, wherea v ailable),ssF CM,ISG,andradiologist-labeledgroundtruth.AppendixBlists theperformanceofthekno wledge-basesystemoneac hsliceprocessed.

PAGE 181

166 6.3.4 Kno wledge-Basedvs.OtherEorts Threeeortsreview edinChapter3hadquan tiableresultsthatcanbecomparedwiththeperformanceofthekno wledge-basedsystem.Sincedieren tdatasets w ereusedineac hoftherespectiv esystems,onlyindirectcomparisonsaremadeb y calculatingthekno wledge-basedsystem'sa v erageperformanceusingtheform ulaof eac hofthesystemsthatw erecomparedandconclusionsthatcanbedra wnfrom theresultsarelimited.Tw oa v eragesarecalculated,a\per-slice"a v erage,whic his simplya v eragesacrossallpatien tsliceslistedinAppendixB,anda\per-v olume"a verage,whic hconsidersthesegmen tationofanen tiretumorv olumewhencalculating thea v erage.Whencalculatingtheseform ulasforthekno wledge-basedsystem,the labels KB and GT willrefertothekno wledge-basedandgroundtruthtumorslice(or v olume)segmen tationsrespectiv ely Ofthesystemsreview edinChapter3,onlyonew orkthatsegmen tedtumor hadquan tiableresults. Thetumorsoft w ometastaticpatien ts,witheigh tand fourslicesrespectiv ely ,w eresegmen tedb yabac k-propagatingnet w ork(BPN)and maxim umlik elihoodclassier(MLC)b y Ozk an,Da w an t,andMaciunas[93].These segmen tationsw erecomparedwiththosegiv enthreeph ysiciansandev aluatedusing asimilarit yindex,denedas: Similarity = S1 T S2 S1 S S2 100 (6.4) where S1 and S2 arethet w oimagesbeingcompared.TheBPNandMLChadana veragesimilarit yindexof53and46respectiv elyagainsttheph ysiciansegmen tations, whilein ter-ph ysiciansimilarit yw asmeasuredat55.Incomparison,tumorsegmen tationsgeneratedb ythekno wledge-basedsystem,hadana v erageper-slicesimilarit yof 57.79anda v erageper-v olumesimilarit yof65.19withgroundtruth(where S 1= KB and S 2= GT ).Theresultsofthekno wledge-basedsystemha v ealev elofsignicance

PAGE 182

167 at =0 : 01,butitshouldbenotedthatthea v erageper-slicesimilarit yindexforthe kno wledge-basedsystemw ascalculatedo v er387patien tslices,whileonly12slices w ereconsideredb ytheabo v emethod. Resultsforthe\iteratedconditionalmodes"(ICM)methodb yJohnston,A tkins, Mac kiewic h,andAnderson[56]insegmen tingMSlesionsw ereev aluatedusingasimilarit yindexbet w eent w oimages S1 and S2 ,denedas: Similarity =2 j S 1 \ S 2 j j S 1 j + j S 2 j (6.5) Lesionsegmen tationcorrespondencebet w eentheirsystemandman uallabelingan a v eragesimilarit yof0 : 505for5v olumesof21-27sliceseac h(appro ximately120slices). Usingtheirsimilarit yindexinev aluatingtumorsegmen tationbet w eenthe KB and GT segmen tations(where S 1= KB and S 2= GT )resultedinana v erageslicev alue of0 : 69(387slices)anda v eragev olumev alueof0 : 78(33v olumes).Bothkno wledgebasedresultsaresignican tat =0 : 01,butsincetheICMmethodsegmen tedMS lesions,thesignicanceisratherlimited. A\percen tagedierence"isusedb yW ells,Grimson,KikinisandJolesz[128], calculatedas: Difference = j S1 X OR S2 j j S 2 j (6.6) Usingthisform ula(with S 1= KB and S 2= GT ),thekno wledge-basedsystem hasana v eragepercen tdierencewithgroundtruthtumorof114%onaper-slice basisand57%per-v olume.Thehigherpercen tagedierencesb ythekno wledge-based systemareconsisten twiththesystem'stendencytoo v erestimatetumorv olumewhen comparedtogroundtruth,asdiscussedabo v e.Theauthors'adaptiv esegmen tation methodhadpercen tagedierencev aluesrangingfrom19-23%againstv eman ual segmen tations,butasnotedinSection3.2.3,onlyasinglesagittalPDslicew as beingev aluatedforarelativ elylimitedtaskofclassicationbet w eenwhiteandgra y

PAGE 183

168 matter,andonlyinareaswherefourofthev eman ualratershadagreemen tin tissuelabels.Nosignican tconclusionscanbedra wnincomparingthet w omethods astheadaptiv esegmen tationproblemism uc hsimplerthanoursandourapproac his expectedtoperformsimilarlyundersuc hconditions.Thisdissertationfocusesonthe problemoftumorsegmen tation,butthekno wledge-basedsystemhasbeenapplied tocompletelysegmen tnormalbraintissuesinpartialv olumesabo v ethev en tricles. ResultsandcomparisonswithkNNsegmen tationscanbefoundin[18]. Lastly ,althoughnoperformancemeasureisgiv en,P annizzo,Stallmey er,and et al. [94]reportthattheirhistogramanalysiscorrectlyfollo w edthegro wth/shrink age ofMSlesionso v ertimeint w elv eofthefourteen(86%)repeatscans.Incomparison, thekno wledge-basedsystemcorrectlytrac k edgroundtruthtumorin22of25(88%) transitions.

PAGE 184

169 CHAPTER7 SUMMAR Y 7.1 SystemSummary Akno wledge-basedsystemthatautomaticallydetects,segmen ts,andlabels glioblastoma-m ultiform etumorinmagneticresonanceimagesofthebrainhasbeen described. Thesystemisanexpertsystemthatiscapableofimageprocessing andm ultispectralanalysis. Theguidanceofthekno wledgebase,ho w ev er,giv es thissystemadditionalpo w erandexibilit yb yallo wingclassicationdecisionsto bemadethroughiterativ e/successiv erenemen twithoutan ysupervisionorreliance onaparticulartrainingset.Thisisincon trasttootherm ultispectraleortswhic h attempttosegmen ttheen tirebrainimageinonestep,basedoneitherstatistical methods[111,119,38,57,128],(un)supervisedclassicationsuc haskNNorssF CM[6,25,26,126,23,114,117],orneuralnet w orks[2,73,93,61]. Qualitativ emodelingw ascompletedforthebraincerebrum,resultingina systemthatcandetectandsegmen tenhancingpathologyinan ytransaxialsliceintersectingthecerebrum.Thekno wledgebasew asbuiltstartingwithageneralset ofheuristicscomparingtheeectsofdieren tpulsesequencesondieren tt ypesof tissues,assho wninSection2.2.Informationabouttheanatomicalstructureofbrain tissuesandthosesurroundingthebrain,suc hastheocularnerv esandm uscles,w ere addedtopro videkno wledgethatisindependen tofan yscanningprotocol,allo wing

PAGE 185

170 suppositionsbasedinfeaturespacetobev eriedinanatomicalorimagespace.This o v erallprocessiscalled\kno wledgeengineering"becausedecisionsm ustbemadeas towhic hkno wledgeismostusefulforaparticulargoal,suc haspathologydetection ortumorsegmen tation,asw ellasho wtoecien tlyimplem en tthekno wledgein toa rule-basedsystem. Thekno wledgebasew asbuiltonsliceswitharelativ elylargethic knessofat least4mm.Thinnerslices,whic hexhibitareducedpartialv olumeeectandallo w bettertissuecon trast,w ouldpermitmoreaccuratesegmen tationofnormalbrain tissuesanddelineationoftumor.Newadv ancesinimagingsoft w are,dev elopedb y ateamattheUniv ersit yofSouthFlorida,ma ypro videimageswithmoredistinct tumorboundaries[77].Thesystemw asdev elopedusingtherelativ erelationshipsof tissuesinfeaturespace,a v oidingdependenceuponspecicfeature-domainv alues.As AppendixAsho ws,an um berofdieren timagingprotocolsw ereusedinacquiring bothv olun teerandpatien tscans,y etextractedkno wledge,suc hasrelativ etissue distributionsinfeaturespace,w asrelativ elyrobust.Gadoliniumhasalsobeenfound tobegenerallyv eryrobustindieren tprotocolsandslicethic knesses[11]. The exten tofthisrobustnesshasnotbeenrigorouslytested,ho w ev er.Shouldacquisition parameterdependencebecomeanissue,giv enalargeenoughtrainingbaseacross m ultipleparameters,kno wledgeengineeringcouldbeusedtoin tegratethec hanges oftissuedistributionsinfeaturespaceunderdieren tscanningparameters,either explicitly(observingcommonlyusedprotocols),orb ylearningthegeneralc hangesin featuredistributionsasthescanningprotocolsaremanipulated.Oncethesec hanges arein tegrated,thekno wledgebasecouldautomaticallyadjusttoaslice'sspecic parameterssincesuc hinformationiseasilyincludedwhenprocessingstarts. Sincethekno wledgeusedinthesystemw asextractedandin tegratedexplicitly(intheformofrules),v ericationmethodsforlimiteddatasets,suc has\lea v e oneout,"cannotbeapplied.Theslicesusedfortraining,ho w ev er,w ereselectedto

PAGE 186

171 encompasstherangeofpossiblecasesforglioblastoma-m ultiform e,includingtumor size,locationwithinthebrain,o v erallenhancemen tinT1andPDspace,andho w distinctthetumorw asfromsurroundingtissue.Thisincreasedthesizeofthetraining set,butonethatw asabletocapturetrendscommonacrossallpatien tswhilestill beingrelativ elysmall.Thisconceptisdemonstratedb ynotingthatmostoftherules usedfortumorsegmen tationw ereoriginallybasedonasev en teenslicetrainingset constructedb ythisauthorin[21]duringinitialdev elopmen tofthetumorsegmentationsystem.Thisoriginaltrainingsetw asthenexpandedtotheonesho wnhere toreco v erlosttumorpixelsinlo w erslices(especiallyforT emplate5and5Lslices), consideradditionalanatomicalstructureswhenremo vingnon-tumorregions,suc has theocularm uscles,andtoadjustthresholdsforT emplate5Lslices.Thissuggests thateectiv ekno wledgehadbeenoriginallyextractedandimpleme n tedin[21]. F urthermore,oneofthesubjectsusedfortrainingw asconsideredextremely dicult,ev enb ytrainedradiologists[84],duetoitsdiuseboundaries,fragmen tation fromradiationandc hemo-therap y ,andthesurgicalresectionofapreviousgro wth. Mostglioblastoma-m ultiform esareunlik elytobenearlyascomplicated,suc hasthe remainingsubjectspresen tedhere,sothekno wledge-basew as,ineect,designedfor moredicultcases.Also,thepatien tv olumesprocessedhadreceiv edv ariousdegrees oftreatmen t,includingsurgery ,radiationandc hemo-therap ybothbeforeandbet w een scans.Y et,despitethec hangesthesetreatmen tscancause,suc hasdem y eli nizationof whitematter,nomodicationstothekno wledge-basedsystemw erenecessary .Other approac hes,lik eneuralnet w orks[2,73,93,61]oran ysupervisedmethodinitially trainedwithaspecicsetofexemplars,couldha v edicultiesindealingwithsligh tly dieren timagingprotocolsandtheeectsoftreatmen t. Asstatedinthein troduction,nomethodofquan titatingtumorv olumesis widelyacceptedandusedinclinicalpractice[83].Anmethodb ytheEasternCooperativ eOncologygroup[34]appro ximatestumorareainthesingleMRslicewiththe

PAGE 187

172 largestcon tiguous,w ell-denedtumor.Thelongesttumordiameterism ultipliedb y itsperpendiculartoyieldanarea.Changesgreaterthan25%intheareaofatumor o v ertimeareused,inconjunctionwithvisualobserv ations,toclassifytumorresponse totreatmen tin tov ecategoriesfromcompleteresponse(nomeasurabletumorleft) toprogression.Thisapproac hdoesnotaddressfulltumorv olume,dependsonthe exactboundaryc hoices,andtheshapeofthetumor[67,34].Byitself,theapproac h canleadtoinaccurategro wth/shrink agedecisions[24]. Th us,thereisaneedforafullyautomaticmethodfortumorv olumemeasuremen t,bothfortrac kingtumorresponsetotherap y ,asw ellastheofplanningfuture treatmen t[69,67,121,24].Thepromisesoftheproposedapproac haredemonstrated b ythesuccessfulperformanceofthesystemontheprocessedslices,bothinpathology detectionandtumorsegmen tation.ThenalKBsegmen tationscomparesw ellwith radiologist-labeled\groundtruth"imagesofthetumor.Thekno wledge-basedsystem alsoperformedw ellagainstthesupervisedkNN,ssF CM,andISGmethodsanddidso withouttheneedfor(m ultiple )h uman-basedR OI'sorpost-processing,whic hmak e kNNclinicallyimpractical.Thekno wledge-basedsystem'sabilit ytoautomatically trac katumor'sgro wth/shrink ageinresponsetotreatmen tasbeenalsodemonstrated. Thekno wledge-basedsystemalsoperformsw ellwhencompared(indirectly)toother systemsreview edinChapter3usingtherespectiv ebenc hmarkform ulasofthosesystems.Otherattemptsatkno wledge-guidedsystemsarebeginningtoappear,suc has thew orksdiscussedinChapter3,andalsosho wsomesuccess. 7.2 F utureW ork Completetissuelabeling(bothnormalandabnormal)inthelo w erslicesother thantumorw asnotofconcernatthispoin t,thoughrulescouldbedev elopedto dosoandisagoodfutureeort.Sincethekno wledge-guidedsystemiscompletely

PAGE 188

173 reproducible,serialmeasuremen tsofbraintissuev olume(usefulforapplicationssuc h asdiagnosis/treatmen tofAlzheim er'spatien ts)m ustonlyconsidersignalv ariations intheMRcoil. W orkb yV aidy anathanin[114]hassho wnanupperlimitonthe v ariationof\constan t"objects(normalbraintissuesofv olun teersubjects)inthe MRcoilusedinthisresearc ho v errepeatacquisitionso v ertime. Otherfuturew orkcouldincludecon tin uedtemplatedev elopmen tfortheremainderofthecerebellumandbrain-stem.Newtumort ypes,especiallythosethat ha v ec haracteristicssimilartotheglioblastoma-m ultiform esstudieshere,shouldbe considered.Asnew erMRIsystemsbecomea v ailable,moreisotropicdata(e.g.,thinnerslices)canbeacquired,asw ellasadditionalfeatures,suc hasdiusionimages, whic hcanbereadilyincludedin tothekno wledgebase. Infact,thisisoneofthe kno wledge-basedsystem'sprimaryadv an tages.Specically ,askno wledgeengineeringisappliedtoadditionalinformationandprocessingtools,usablekno wledgema y beextractedandin tegratedtoimpro v ethesystem'sperformance.F orexample,as describedinAppendixC,edgedetectioninformationma ybeusedaskno wledgein settingthenalthresholdinStageFiv e. Preliminaryresultsareencouragingfor tumorcaseswithdistinctboundaries,butdicultiesarefoundwithmorediuse tumors. Thekno wledge-basedapproac hisdesigned,ho w ev er,suc hthatoncethe tec hniqueisrened,itma ybefullyin tegratedin tothesystem.Akno wledgebase alsoallo wsstraigh tforw ardexpansionin toadditionaltumort ypes,asw ellasother brainabnormalities,suc hasMSlesionsorpossiblyheadtrauma.Outsidethescope ofthisresearc h,acommondatabasewithbiopseypro v edandlocatedpathology w ouldpro v eusefulintheev aluationandcomparisonofthev ariousMRsegmen tation methodsdiscussedhere. Inconclusion,thesystempresen tedhereisanexpertsystemwithcapabilitiesof imageprocessingandm ultispectralanalysisthatcaneectiv elysegmen tglioblastomam ultiform etumorsusingrulesandheuristicsthatareindependen tofaparticular

PAGE 189

174 scanningprotocol.Itsunsupervisednatureremo v estheneedforh umansupervision andthein ter/in traobserv erv ariabilit yitbrings,andhasthepoten tialofbeingausefultoolfordetectingpathology ,segmen tingtumorfortherap yplanning,andtrac king tumorresponsetotherap y .Thekno wledge-basedparadigmallo wseasyin tegration ofnewdomaininformationandprocessingtoolsin totheexistingsystem. Finally thekno wledge-basedapproac hpromisesstraigh tforw ardexpansionin toothert ypes ofpathologyandMRdata.

PAGE 190

175 LISTOFREFERENCES [1]J.Alirezaie,M.E.Jernigan,andC.Nahmias.Neural-net w orkbasedsegmen tationofmagneticresonanceimagesofthebrain. IEEET r ansactionsonNucle ar Scienc e ,44(2):194{198,April1997. [2]S.C.Amartur,D.Piriano,andY.T ak efuji.Optimizationneuralnet w orksfor thesegmen tationofmagneticresonanceimages. IEEETMI ,11(2):215{221, June1992. [3]A.Bensaid. Impr ove dF uzzyClusteringforPatternR e c o gnitionwithApplic ationstoImageSe gmentation .PhDthesis,Univ ersit yofSouthFlorida,1994. [4]A.M.Bensaid,L.O.Hall,J.C.Bezdek,andL.P .Clark e.P artiallysupervised clusteringforimagesegmen tation. PatternR e c o gnition ,29(5):859{871,1996. [5]J.C.Bezdek,L.O.Hall,M.C.Clark,andD.B.Goldgof. Segmen tingmedical imageswithfuzzymodels:Anupdate.InD.Dubois,H.Prade,andR.Y ager, editors, F uzzyInformationEngine ering ,pages69{92.JohnWileyandSons: NewY ork,1997. [6]J.C.Bezdek,L.O.Hall,andL.P .Clark e.ReviewofMRimagesegmen tation tec hniquesusingpatternrecognition. Me dic alPhysics ,20(4):1033{1048,1993. [7]J.C.BezdekandS.K.P al. F uzzyMo delsforPatternR e c o gnition .IEEEPress: Piscata w a y ,NJ,1992. [8]M.Bomans,K.H.H ohne,U.Tiede,andM.Riemer.3Dsegmen tationofMR imagesoftheheadfor3Ddispla y IEEETMI ,9:177{183,1990. [9]P .Bottomley ,T.F oster,R.Argersinger,andL.Pfeier.Areviewofnormal tissueh ydrogenNMRrelaxationtimesandrelaxationmec hanismsfrom1-100 MHz:Dependencyontissuet ype,NMRfrequency ,temperature,species,excisionandage. Me dic alPhysics ,11:425{448,1984. [10]M.E.Brandt,T.P .Bohan,L.A.Kramer,and etal .EstimationofCSF,white andgra ymatterv olumesinh ydrocephalicc hildrenusingfuzzyclusteringof MRimages. Computerize dMe dic alImagingandGr aphics ,18(1):25{34,1994. [11]R.BronenandG.Sze.Magneticresonanceimagingcon trastagen ts:Theoryand applicationtothecen tralnerv oussystem. JournalofNeur osur gery ,73:820{839, 1990.

PAGE 191

176 [12]M.E.Brummer.Optimizedin tensit ythresholdforv olumetricanalysisofmagneticresonanceimagingdata.In VisualizationinBiome dic alComputing1992. Pr o c e e dingsSPIE1808 ,pages299{310,1992. [13]R.L.Cannon,J.V.Da v e,andJ.C.Bezdek. Ecien timplem en tationofthe fuzzyc-meanclusteringalgorithms. IEEET r ansactionsonPatternA nalysis andMachineIntelligenc e ,8(2):248{255,1986. [14]C.W.Chang,G.R.Hillman,H.Ying,and etal .Segmen tationofratbrainMR imagesusingah ybridfuzzysystem.In Pr o c e e dingsofthe19941stInternational JointConfer enc eofNAFIPS/IFIS/NASA ,pages55{59.NAFIPS/IFIS/NASA, 1994.SanAn tonio,TX. [15]C.W.Chang,G.R.Hillman,H.Ying,and etal At w o-stageh umanbrain MRIsegmen tationsc hemeusingfuzzylogic. In Pr o c e e dingsoftheInternationalJointConfer enc eoftheF ourthIEEEInternationalConfer enc eonF uzzy SystemsandtheSe c ondInternationalF uzzyEngine eringSymp osium ,Marc h 1995.Y ok ohama,Japan. [16]R.T.ChinandC.R.Dy er. Model-basedrecognitioninrobotvision. A CM ComputingSurveys ,18(1):67{108,1986. [17]H.S.Choi,D.R.Ha ynor,andY.Kim.P artialv olumeclassicationofm ultic hannelmagneticresonanceimages-amixelmode. IEEETMI ,10(3):395{407, 1991. [18]MatthewClark.Segmen tingMRIv olumesofthebrainwithkno wledge-based clustering.Master'sthesis,Univ ersit yofSouthFlorida,1994. [19]M.C.Clark,L.O.Hall,andD.B.Goldgof.Usingfuzzyinformationinkno wledge guidedsegmen tationofbraintumors.In F uzzyL o gicinAIWorkshopatIJCAI 1995 ,1995.Mon treal,CA. [20]M.C.Clark,L.O.Hall,D.B.Goldgof,and etal .MRIsegmen tationusingfuzzy clusteringtec hniques:In tegratingkno wledge. IEEEEngine eringinMe dicine andBiolo gy ,13(5):730{742,1994. [21]M.C.Clark,L.O.Hall,D.B.Goldgof,R.P .V elth uizen,andM.S.Silbiger. Automatictumorsegmen tationusingkno wledge-basedtec hniques.Submitted toIEEE:TMI. [22]M.C.Clark,L.O.Hall,C.Li,andD.B.Goldgof. Kno wledgebased(re)clustering.In Pr o c e e dingsofthe12thIAPRInternationalConfer enc eonPatternR e c o gnition ,pages245{250,1994.Jerusalem,Israel. [23]L.P .Clark e,R.P .V elth uizen,M.Camac ho,J.Heine,M.V a ydianathan,L.O. Hall,R.Thatc her,andM.Silbiger.MRIsegmen tation:Methodsandapplications. MagneticR esonanc eImaging ,13(3):343{368,1995.

PAGE 192

177 [24]L.P .Clark e,R.P .V elth uizen,M.C.Clark,G.Ga viria,L.O.Hall,D.Goldgof,and etal .MRImeasuremen tofbraintumorresponse:Comparisonofvisualmetric andautomaticsegmen tation.SubmittedtoMagneticResonanceImaging,June 1997. [25]L.P .Clark e,R.P .V elth uizen,L.O.Hall,J.C.Bezdek,A.M.Bensaid,andM.L. Silbiger.Comparisonofsupervisedpatternrecognitiontec hniquesandunsupervisedmethodsforMRIsegmen tation.In SPIE:Me dic alImagingVI:Image Pr o c essing ,v olume1652,pages668{677,1992. [26]L.P .Clark e,R.P .V elth uizen,S.Ph uphanic h,J.D.Sc hellen berg,J.A.Arrington, andM.Silbiger. MRI:Stabilit yofthreesupervisedsegmen tationtec hniques. MagneticR esonanc eImaging ,11:95{106,1993. [27]G.Cohen,B.Andreasen,R.Alliger,and etal Segmen tationtec hniquesfor theclassicationofbraintissueusingmagneticresonanceimaging. Psychiatry R ese ar ch:Neur oimaging ,45:33{51,1992. [28]B.V.Dasarth y Ne ar estNeighb or(NN)Norms:NNPatternClassic ationT e chniques .IEEEComputerSociet yPress,LosAlamitos,Ca.,1991. [29]M.C.deOliv eiraandR.I.Kitney .T extureanalysisfordiscriminationoftissues inMRIdata.In ComputersinCar diolo gy ,pages481{484,1992. [30]S.Dellepiane. Imagesegmen tation:Errors,sensitivit y ,anduncertain t y In Pr o c e e dingsofthe13thIEEEEMBSo ciety ,v olume13,pages253{254,1991. [31]S.Dellepiane,G.V en turi,andG.V ernazza.Afuzzymodelfortheprocessing andrecognitionofMRpathologicalimages.In IPMI1991 ,pages444{457,1991. [32]ThomasC.F arrar. A nIntr o ductiontoPulseNMRSp e ctr osc opy F arragut Press,1987. [33]G.BonnieF eldman.Braintumorfactsandgures.TheBrainT umorSociet y -h ttp://www.th ts.org/btfacts.h tm,July31997. [34]L.F eun.Double-blindrandomizedtrialofthean ti-progestationalagen tmifepristoneinthetreatmen tofunresectablemeningioma,phaseiii.T ec hnicalReport SW OG-9005,Univ ersit yofSouthFlorida,T ampa,Fl.,South w estOncology Group,1995. [35]L.M.Fletc her-Heath,J.B.Barsotti,andJ.P .Hornak.Am ultispectralanalysis ofbraintissues. MagneticR esonanc einMe dicine ,29:623{630,1993. [36]K.S.F uandJ.K.Mui.Asurv eyonimagesegmen tation. PatternR e c o gnition 13:3{16,1981. [37]R.Gallo w a y ,R.J.Maciunas,andA.F ailinger.F actorsaectingperceiv edtumor v olumesinmagneticresonanceimaging. A nnalsofBiome dic alEngine ering 21:367{375,1993.

PAGE 193

178 [38]G.Gerig,J.Martin,R.Kikinis,and etal .Automatingsegmen tationofdualec hoMRheaddata. In The12thInternationalConfer enc eofInformation Pr o c essinginMe dic alImaging(IPMI1991) ,1991. [39]G.Gerig,J.Martin,R.Kikinis,O.Kubler,M.Shen ton,andF.A.Jolesz.Unsupervisedtissuet ypesegmen tationof3Ddual-ec hoMRheaddata. Imageand VisionComputing ,10:349{360,1992. [40]J.GiarratanoandG.Riley Exp ertSystems: PrinciplesandPr o gr amming Boston:PWSPublishing,secondedition,1994. [41]P .Gibbs,D.L.Buc kley ,S.J.Blac kband,andA.Horsman.T umorv olumedeterminationfromMRimagesb ymorphologicalsegmen tation. PhysicsinMe dicine andBiolo gy ,41(11):2437{2446,No v em ber1996. [42]L.GongandC.A.Kulik o wski.Automaticsegmen tationofbrainimages:Selectionofregionextractionmethods. In SPIEV ol.1450Biome dic alImage Pr o c essingII ,pages144{153.SPIE,1991. [43]D.E.GustafsonandW.C.Kessel. F uzzyclusteringwithafuzzyco v ariance matrix.In IEEECDC ,pages761{766,1979.SanDiego,CA,Jan10-12. [44]D.E.Haines. Neur o anatomy:AnA tlasofStructur es,Se ctions,andSystems WilliamsandWilkins,fourthedition,1995. [45]L.O.Hall,A.M.Bensaid,L.P .Clark e,and etal .Acomparisonofneuralnet w ork andfuzzyclusteringtec hniquesinsegmen tingmagneticresonanceimagesofthe brain. IEEET r ansactionsonNeur alNetworks ,3(5):672{682,1992. [46]R.M.Haralic k,K.Shanm ugam,andI.Dinstein.T exturalfeaturesforimage classication. IEEET r ansactionsonSystems,Man,andCyb ernetics ,3(6):610{ 621,1973. [47]J.A.Hartigan. ClusteringA lgorithms SanDiego: AcademicPress,second edition,1975. [48]D.H.HearnandM.P .Bak er. ComputerGr aphics .Pren tice-Hall,Inc.,1986. [49]R.Hendric kandE.Haac k e.Basicph ysicsofMRcon trastagen tsandmaximizationofimagecon trast. JMRI ,3(1):137{148,1993. [50]G.R.Hillman,C.Chang,H.Ying,and etal .AutomaticsystemforbrainMRI analysisusingano v elcom binationoffuzzyrule-basedandautomaticclustering tec hniques.In Me dic alImaging1995:ImagePr o c essing ,pages16{25.SPIE, F ebruary1995.SanDiego,CA. [51]M.Holden,E.Steen,andA.Lunderv old. Segmen tationandvisualizationof brainlesionsinm ultispectralmagneticresonanceimages. Computerize dMe dic al ImagingandGr aphics ,19(2):171{183,1995. [52]B.Horn. R ob otVision .NewY ork:McGra w-Hill,1986.

PAGE 194

179 [53]T.J.Hyman,R.J.Kurland,G.C.Levy ,andJ.D.Shoop. Characterizationof normalbraintissueusingsev encalculatedMRIparametersandastatistical analysissystem. MagneticR esonanc einMe dicine ,11:22{34,1989. [54]A.K.Jain. F undamentalsofDigitalImagePr o c essing Englew oodClis,NJ: Pren ticeHall,1989. [55]R.Jain,R.Kasturi,andB.Sc h unc k. MachineVision .McGra w-Hill,Inc.,1995. [56]B.Johnston,M.S.A tkins,B.Mac kiewic h,andM.Anderson. Segmen tation ofm ultiplesclerosisinin tensit ycorrectedm ultispectralMRI. IEEETMI 15(2):154{169,April1996. [57]M.Kam ber,R.Shingal,D.Collins,G.F rancis,andA.Ev ans.Model-based3D segmen tationofm ultiplesclerosislesionsinmagneticresonancebrainimages. IEEETMI ,14(3):442{453,1995. [58]N.Karssemeijer.Astatisticalmethodforautomaticlabelingoftissuesinmedicalimages. MachineVisionandApplic ations ,3:75{86,1990. [59]N.Keh tarna v az,M.Ch ung,L.A.Ha yman,andIIIR.E.W endt.Magneticresonanceimagesegmen tationb ycon textualfuzzyclustering. JournalofIntelligent andF uzzySystems ,1:295{305,1994. [60]R.Kikinis,M.Shen ton,G.Gerig,and etal .Routinequan titativ eanalysisof brainandcerebrospinaluidspaceswithMRimaging. JMRI ,2:619{629,1992. [61]E.R.Kisc hell,N.Keh tarna v az,G.R.Hillman,H.Levin,M.Lilly ,andT.A. Ken t.Classicationofbraincompartmen tsandheadinjurylesionsb yneural net w orksappliedtoMRI. Neur or adiolo gy ,37:535{541,1995. [62]L.Kjr,P .Ring,C.Thomsen,andO.Henriksen.T extureanalysisinquan titativ eMRimaging. A ctaR adiolo gic a ,36(2):127{135,1995. [63]M.KobashiandL.G.Shapiro.Kno wledge-basedorganiden ticationfromCT images. PatternR e c o gnition ,28(4):475{491,1995. [64]MasaharuKobashi. Kno wledge-basedorganiden ticationfromCTimages. Master'sthesis,Univ ersit yofW ashington,1992. [65]M.I.Kohn,N.K.T anna,G.T.Herman,and etal .AnalysisofbrainandcerebrospinaluidspacesinMRimaging. R adiolo gy ,178:115{122,1991. [66]A.Kundu. Localsegmen tationofbiomedicalimages. Computerize dMe dic al ImagingandGr aphics ,14:173{183,1990. [67]N.LaperrireandM.Berstein.Radiotherap yforbraintumors. CA-ACanc er JournalforClinicians ,4:96{108,1994. [68]R.Leah y ,T.Hebert,andR.Lee.ApplicationsofMark o vrandomeldmodels inmedicalimaging.In Pr o cIPMI ,v olume11,pages1{14,1989.

PAGE 195

180 [69]N.LeedsandE.Jac kson. Curren timagingtec hniquesfortheev aluationof brainneoplasms. Curr entScienc e ,6:254{261,1994. [70]C.Li,D.B.Goldgof,andL.O.Hall.Automaticsegmen tationandtissuelabeling ofMRbrainimages. IEEETMI ,12(4):740{750,Decem ber1993. [71]Ch unlinLi. Kno wledgebasedclassicationandtissuelabelingofmagnetic resonanceimagesofthebrain. Master'sthesis,Univ ersit yofSouthFlorida, 1993. [72]H.Li,R.Deklerc k,B.DeCuyper,A.Herman us,E.Nyssen,andJ.Cornelis. ObjectrecognitioninbrainCT-scans: Kno wledge-basedfusionofdatafrom m ultiplefeatureextractors. IEEETMI ,14(2):212{229,June1995. [73]X.Li,S.Bhide,andM.Kabuk a.LabelingofMRbrainimagesusingboolean neuralnet w ork. IEEETMI ,15(2):628{638,1996. [74]K.O.LimandA.Perbaum. Segmen tationofMRbrainimagesin tocerebrospinaluidspaces,white,andgra ymatter. JournalofComputerAssiste d T omo gr aphy ,13:588{593,1989. [75]J.S.Lin,K.S.Cheng,andC.W.Mao.Multispectralmagneticresonanceimages segmen tationusingfuzzyhopeldneuralnet w ork. InternationalJournalof Biome dic alComputing ,42(3):205{214,August1996. [76]J.S.Lin,K.S.Cheng,andC.W.Mao.Segmen tationofm ultispectralmagnetic resonanceimageusingpenalizedfuzzycompetitiv elearningnet w ork. Comput Biome dR es ,29(4):314{326,August1996. [77]W.Link.Computersoft w arema yhelpgh tcancer. TheOr acle ,page3,October 21997. [78]RobertB.Lufkin. TheMRIManual .Y earBookMedicalPublishers,Inc.,1990. [79]G.F.LugerandW.A.Stubbleeld. A rticialIntelligenc e:Structur eandStr ategiesforComplexPr oblemSolving .TheBenjamin/Cumm ingsPublishingCo., Inc.,secondedition,1993. [80]A.Lunderv oldandG.Storvik. Segmen tationofbrainparenc h ymaand cerebrospinaluidinm ultispectralmagneticresonanceimages. IEEETMI 14(2):339{349,June1995. [81]W.Menhardt. IconicfuzzysetsforMRimagesegmen tation. Presen tedat NA TO/ASIonMedicalImaging,1988. [82]W.MenhardtandK.Sc hmidt.Computervisiononmagneticresonanceimages. PatternR e c o gnitionL etters ,8(2):73{85,Septem ber1988. [83]R.Murtagh,S.Ph uphanic h,N.Imam,L.Clark e,M.V aidy anathan,and et.al. No v elmethodsofev aluatingthegro wthresponsepatternsoftreatedbrain tumors. Canc erContr ol ,pages293{299,1995.

PAGE 196

181 [84]F.R.Murtaugh.DiscussionsheldwithDr.F.ReedMurtaugh,M.D.,Dept.of Radiology ,Univ ersit yofSouthFlorida,April1997. [85]F.R.Murtaugh.DiscussionsheldwithDr.F.ReedMurtaugh,M.D.,Dept.of Radiology ,Univ ersit yofSouthFlorida,October,11997. [86]F.R.Murtaugh.DiscussionsheldwithDr.F.ReedMurtaugh,M.D.,Dept.of Radiology ,Univ ersit yofSouthFlorida,October,221997. [87]A.Namasiv a y am. Segmen tationofmagneticresonanceimagesofthebrain usingfuzzyrules.Master'sthesis,Univ ersit yofSouthFlorida,1996. [88]A.Namasiv a y amandL.O.Hall.Useoffuzzyrulesinclassicationofnormal h umanbraintissues.In Pr o c3IntSympUnc ertMo delA nalA nnuConfNorth A merF uzzyInfPr o c essSo c(ISUMA-NAFIPS1995) ,pages157{162,1995. CollegeP ark,MD. [89]A.Namasiv a y amandL.O.Hall.In tegratingfuzzyrulesin tothefast,robustsegmen tationofmagneticresonanceimages.In NewF r ontiersinF uzzyL o gicand SoftComputingBiennialConfer enc eoftheNorthA meric anF uzzyInformation Pr o c essingSo ciety-NAFIPS1996 ,pages23{27,1996.Piscata w a y ,NJ. [90]A.M.NazifandM.D.Levine.Lo wlev elimagesegmen tation:Anexpertsystem. IEEET r ansactionsonPatternA nalysisandMachineIntelligenc e ,6(5):555{ 577,1984. [91]R.A.No v ellineandL.F.Squire. LivingA natomy .HanleyandBelfus,1987. [92]R.J.Ott,M.A.Flo w er,J.W.Babic h,andP .K.Marsden. Theph ysicsofradioisotopeimaging.InS.W ebb,editor, ThePhysicsofMe dic alImaging ,pages 142{318.InstituteofPh ysicsandPublishing:BristolandPhiladelphia,1988. [93]M. Ozk an,B.M.Da w an t,andR.J.Maciunas.Neural-net w ork-basedsegmentationofm ulti-m odalmedicalimages:Acomparativ eandprospectiv estudy IEEETMI ,12(3):534{545,Septem ber1993. [94]F.P annizzo,M.J.Stallmey er,J.F riedman,R.J.Jennis,and etal .Quan titativ eMRIstudiesforassessmen tofm ultiplesclerosis. MagneticR esonanc ein Me dicine ,24:90{99,1992. [95]R.R.Priceand etal .Qualit yassurancemethodsandphan tomsformagnetic resonanceimaging:ReportofAAPMn uclearmagneticresonanceTaskGroup No.1. Me dic alPhysics ,17(2):287{295,1990. [96]U.RaandF.D.Newman.Automatedlesiondetectionandlesionquan titation inMRimagesusingautoassociativ ememory Me dic alPhysics ,19:71{77,1992. [97]S.P .Ra y a.Lo wlev elsegmen tationof3-Dmagneticresonancebrainimages| arule-basedsystem. IEEETMI ,9(3):327{337,1990.

PAGE 197

182 [98]R.Reiter.Alogicfordefaultreasoning. A rticialIntelligenc e ,13(1and2):81{ 132,1980. [99]G.Riley .V ersion4.3CLIPSreferenceman ual.T ec hnicalReportJSC-22948, ArticialIn telligenceSection,LyndonB.JohnsonSpaceCen ter,1989. [100]E.RubinandJ.L.F arber. Patholo gy .Philadelphia:J.P .LippincottCompan y 1988. [101]L.R.Sc had,S.Bl umi,andI.Zuna.MRtissuec haracterizationofin tracranial tumorb ymeansoftextureanalysis. MagneticR esonanc eImaging ,11(6):889{ 896,1993. [102]H.N.Sc hnitzleinandF.ReedMurtaugh. ImagingA natomyoftheHe ad andSpine: APhoto gr aphicColorA tlasofMRI,CT,Gr oss,andMicr osc opicA natomyinAxial,Cor onal,andSagittalPlanes Baltimore: Urban &Sc h w arzen berg,secondedition,1990. [103]J.Serra. ImageA nalysisandMathematic alMorpholo gy .London;NewY ork: AcademicPress,1982. [104]M.Sonk a,S.K.T adik onda,andS.M.Collins.Kno wledge-basedin terpretation ofMRbrainimages. IEEETMI ,15(4):443{452,August1996. [105]V.SridharandM.N.Murt y .Akno wledge-basedclusteringalgorithm. Pattern R e c o gnitionL etters ,12(9):511{517,Septem ber1991. [106]Da vidD.StarkandJr.WilliamG.Bradley MagneticR esonanc eImaging, Se c ondEd.,V olumeOne .Mosb yY earBook,1992. [107]H.SuzukiandJ.T orik a w a.Automaticsegmen tationofheadMRIimagesb y kno wledge-guidedthresholding. Computerize dMe dic alImagingandGr aphics 15(4):233{240,1991. [108]S.L.T animoto. TheElementsofA rticialIntelligenc e .NewY ork:Computer SciencePress,1990. [109]C.W.T aoandW.EThompson. Afuzzyif-thenapproac htoedgedetection. In 1993IEEEInternationalConfer enc eonF uzzySystems ,pages1356{1360. IEEE,1993. [110]T.T axtandA.Lunderv old. Multispectralanalysisofthebraininmagnetic resonanceimaging.In IEEEWorkshoponBiome dic alImageA nalysis ,pages 33{42,1994.LosAlamitos,CA,USA. [111]T.T axtandA.Lunderv old.Multispectralanalysisofthebrainusingmagnetic resonanceimaging. IEEETMI ,13(3):470{481,Septem ber1994. [112]T.T axt,A.Lunderv old,B.F uglaas,H.Lien,andV.Abeler. Multispectral analysisofuterinecorpustumorsinmagneticresonanceimaging. Magnetic R esonanc einMe dicine ,23:55{76,1992.

PAGE 198

183 [113]C.Tsai,B.S.Manjunath,andR.Jagadeesan.Automatedsegmen tationofbrain MRimages. PatternR e c o gnition ,28(12):1825{1837,1995. [114]M.V aidy anathan,L.P .Clark e,C.Heidman,R.P .V elth uizen,andL.O.Hall. Normalbrainv olumemeasuremen tusingm ultispectralMRIsegmen tation. MagneticR esonanc eImaging ,15(1):87{97,1997. [115]M.V aidy anathan,L.P .Clark e,R.P .V elth uizen,S.Ph uphanic h,A.M.Bensaid, L.O.Hall,J.C.Bezdek,H.Green berg,A.T rotti,andM.L.Silbiger.ComparisonofsupervisedMRIsegmen tationmethodsfortumorv olumedetermination duringtherap y MagneticR esonanc eImaging ,13(5):719{728,1995. [116]M.V aidy anathan,R.P .V elth uizen,L.P .Clark e,andL.O.Hall.Quan titationof braintumorinMRIfortreatmen tplanning.In Pr o c e e dingsofthe16thA nnual InternationalConfer enc eoftheIEEEEngine eringinMe dicineandBiolo gy So ciety ,v olume16,pages555{556,1994. [117]M.V aidy anathan,R.P .V elth uizen,P .V en ugopal,andL.P .Clark e. T umor v olumemeasuremen tsusingsupervisedandsemi-supervisedMRIsegmen tation methods.In A rticialNeur alNetworksinEngine ering-Pr o c e e dings(ANNIE 1994) ,v olume4,pages629{637,1994. [118]A.J.V ander,J.H.Sherman,andD.S.Luciano. HumanPhysiolo gy:TheMe chanismsofBo dyF unction .McGra w-HillPublishingCo.,fthedition,1990. [119]M.W.V annier,R.L.Buttereld,D.Jordan,and etal .Multispectralanalysisof magneticresonanceimages. R adiolo gy ,154(1):221{224,Jan uary1985. [120]M.W.V annier,C.M.Speidel,andD.L.Ric kmans.Magneticresonanceimaging m ultispectraltissueclassication. NewsPhysiolSci ,3:148{154,August1988. [121]R.V elth uizen,L.O.Hall,andL.P .Clark e.Unsupervisedfuzzysegmen tation of3Dmagneticresonancebrainimages.In Pr o c e e dingsoftheIS&TSPIE1993 InternationalSymp osiumonEle ctr onicImages:Scienc e&T e chnolo gy ,v olume 1905,pages627{635,1993.SanJose,CA,Jan31-F eb4. [122]R.P .V elth uizen. Lettertotheeditor:Quan titativ eanalysisofbraintissues withMRimaging. R adiolo gy ,183(3):876{877,1992. [123]R.P .V elth uizen.V alidit yguidedclusteringforbraintumorsegmen tation.In A nnualInternationalConfer enc eoftheIEEEEMB-Pr o c e e dings ,v olume17, pages413{414,1995. [124]R.P .V elth uizen. F e atur eExtr actionwithGeneticA lgorithmsforF uzzyClustering .PhDthesis,Univ ersit yofSouthFlorida,1996. [125]R.P .V elth uizenandL.P .Clark e. Anin terfaceforv alidationofMRimage segmen tations.In Pr o c e e dingsofthe16thA nnualInternationalConfer enc eof theIEEEEngine eringinMe dicineandBiolo gySo ciety ,pages547{548,1994.

PAGE 199

184 [126]R.P .V elth uizen,S.Ph uphanic h,L.P .Clark e,L.O.Hall,and etal .Unsupervised tumorv olumemeasuremen tusingmagneticresonancebrainimages. JMRI 5(5):594{605,1995. [127]S.Vinitski,C.Gonzalez,C.Burnett,S.Seshagiri,and etal .Tissuesegmen tation b yhighresolutionMRI:Impro v edaccuracyandstabilit y .In Pr o c e e dingsofthe 16thA nnualInternationalConfer enc eoftheIEEEEngine eringinMe dicine andBiolo gySo ciety ,v olume16,pages557{558,1994. [128]IIIW.M.W ells,W.E.Grimson,R.Kikinis,andF.A.Jolesz.Adaptiv esegmentationofMRIdata. IEEETMI ,15(4):429{441,August1996. [129]C.S.W onandH.Derin.Maxim umlik elihoodestimationofgaussianmark o v randomeldparameters.In Pr o cIEEEICASSP ,pages1040{1043,1990. [130]Z.W uandR.Leah y .Agraphtheoreticapproac htosegmen tationofMRimages. In SPIEV ol.1450Biome dic alImagePr o c essingII ,pages120{132,1991. [131]Z.W uandR.Leah y .Imagesegmen tationviaedgecon tournding:Agraph theoreticapproac h.In IEEEComputerVisionandPatternR e c o gnition ,pages 613{619,1992. [132]Y.Zh uandH.Y an.Computerizedtumorboundarydetectionusingahopeld neuralnet w ork. IEEETMI ,16(2):55{67,F ebruary1997.

PAGE 200

185 APPENDICES

PAGE 201

186 APPENDIXA MRD A T ASETPR OTOCOLS Allv olun teersubjectsliceshadaeld-of-view(F O V)of220mm(pixelsize 0.86mmandimagesize256x256pixels),whilepatien tsliceshadaeld-of-viewof either220mmor240mm(pixelsize0.94mmandimagesize256x256pixels).Allslices ha v eT1-w eigh ted(spinec ho),PD-w eigh ted:(fastspinec ho),andT2-w eigh ted:(fast spinec ho)featureimages.Imagesw ereacquiredusingina1.5T eslaGeneralElectric imagingcoil. TheparticularTR/TEv aluesandslicethic knessesarelistedbelo w. Signaluniformit yw asmeasuredaccordingtoAAPMstandards[95],withacylindrical phan tomwithadiameterof8inc heswhic hw asimagedwithaeld-of-viewof270 mm.T omeasurethew orst-casenon-uniformit y ,nosmoothingw asapplied. Nonuniformit yw asmeasuredforeac htransaxialplane,andresultedinv aluesbet w een 89%and94%forallimagesequences.Nogradien tsinsignalin tensit yw ereobserv ed inthedatasets,norw asan ywithinslicenon-uniformit y .Allimagingw asperformed post-con trast,withgen tlerestrain tsplacedontheheadtoprev en tmo v eme n t,a v oiding an yregistrationproblems. TheMRscannerpro vides12-bitdatawhic hw asused withoutfurtherscaling.Allsubjectswithgadolinium(Magnevist)enhancemen thad aconcen trationof0.1mmol/kg.

PAGE 202

APPENDIXA(Con tin ued) 187 ScanningP arametersofV olun teerSubjects Subject Scan TR/TE(ms)V alue Thic k/Gap F O V T1 PD T2 (mm/m m ) V1 Base 500/27 3000/16 3000/96 4/1 220 V2 Base 500/27 3000/16 3000/96 4/1 220 V3 Base 500/27 3000/16 3000/96 4/1 220 V4a Base 500/27 3000/16 3000/96 4/0 220 V4b Base 500/27 3000/16 3000/96 4/1 220 V4c Base 500/27 3000/17 3000/102.2 5/0 220 R1 500/27 3000/17 3000/102.2 5/0 220 R2 500/27 3000/17 3000/102.2 5/0 220 V5 Base 500/27 3000/17 3000/102.2 5/0 220 V6 Base 500/27 3000/16 3000/96 4/1 220 R1 500/27 3000/16 3000/96 4/0 220 R2 500/27 3000/17 3000/102.2 5/0 220 R3 500/27 3000/17 3000/102.2 5/0 220 R4 500/27 3000/17 3000/102.2 5/0 220 V7 Base 500/27 3000/16 3000/96 4/0 220 R1 500/27 3000/17 3000/102.2 5/0 220 R2 500/27 3000/17 3000/102.2 4/1 220 V8 Base 500/27 3000/17 3000/102.2 5/0 220 V9 Base 500/27 3000/17 3000/102.2 5/0 220 V10 Base 500/27 3000/16 3000/96 4/1 220 R1 500/27 3000/17 3000/102.2 5/0 220 V11 Base 500/27 3000/17 3000/102.2 5/0 220 V12 Base 500/27 3000/17 3000/102.2 5/0 220 V13 Base 500/27 3000/17 3000/102.2 5/0 220 R1 500/27 3000/17 3000/102.2 5/0 220 R2 500/27 3000/17 3000/102.2 5/0 220 GN1 Base 600/27 3000/17 3000/102.2 5/0 220 GN2 Base 500/27 3000/17 3000/102.2 5/0 220 GN3 Base 650/11 3000/17 3000/102.2 5/0 220 GN4 Base 650/11 3000/17 3000/102.2 5/0 220

PAGE 203

APPENDIXA(Con tin ued) 188 ScanningP arametersofP atien tSubjects Subject Scan TR/TE(ms)V alue Thic k/Gap F O V T1 PD T2 (mm/m m ) P1 Base 650/11 4000/17 4000/102 5/0 240 R1 650/11 4000/17 4000/102 5/0 240 R2 650/11 4000/17 4000/102 5/0 240 R3 650/11 4000/17 4000/102 5/0 240 R4 650/11 4000/17 4000/102 5/0 240 P2 Base 650/11 4000/17 4000/102 5/0 240 R1 650/11 4000/17 4000/102 5/0 240 R2 650/11 4000/17 4000/102 5/0 240 P3 Base 650/11 4000/17 4000/102 5/0 240 R1 650/11 4000/17 4000/102 5/0 240 P4 Base 650/11 4000/17 4000/102 5/0 240 R1 650/11 4000/17 4000/102 5/0 240 R2 650/11 4000/17 4000/102 5/0 240 P5 Base 500/27 3000/17 3000/102.2 5/0 220 R1 650/11 3000/17 3000/102.2 5/0 220 R2 650/11 3000/17 3000/102.2 5/0 220 R3 650/11 3000/17 3000/102.2 5/0 220 R4 650/11 3000/17 3000/102.2 5/0 220 P6 Base 650/11 3000/17 3000/102.2 5/0 220 R1 650/11 3000/17 3000/102.2 5/0 220 R2 650/11 3000/17 3000/102.2 5/0 220 R3 650/11 3000/17 3000/102.2 5/0 220 R4 650/11 3000/17 3000/102.2 5/0 220 R5 650/11 3000/17 3000/102.2 5/0 220 R6 650/11 3000/17 3000/102.2 5/0 220 R7 650/11 3000/17 3000/102.2 5/0 220 R8 650/11 3000/17 3000/102.2 5/0 220 R9 650/11 3000/17 3000/102.2 5/0 220 R10 650/11 3000/17 3000/102.2 5/0 220 R11 650/11 3000/17 3000/102.2 5/0 220 R12 650/11 3000/17 3000/102.2 5/0 220 P7 Base 650/11 4000/17 4000/102 5/0 240 P8 Base 650/11 4000/17 4000/102 5/0 240

PAGE 204

189 APPENDIXB TUMORSEGMENT A TIONRESUL TS Aslice-b y-slicelistofthekno wledge-basedsystem'sperformanceonpatien t subjectsaregiv en. Theinitial/cen tersliceforaparticularv olumeisindicatedb y a\C,"intheslicecolumn(Sl.)whilean\L"indicatestheslicew asT emplate5L. T rainingslicesaremark edb yav alueinthe\T r." column,where\All"indicates aslicew asusedastraininginallprocessingstages,bothpathologydetectionand tumorsegmen tation.\P"indicatestheslicew asusedastrainingonlyforpathology detection.\A T"indicatestheslicew asusedastrainingforalltumorsegmen tation stages,while\TS5"indicatestheslicew asusedonlyforStageFiv e(thenalT1 threshold). Slicesmark edwithan\N"w ereconsiderednormalb ythepathology detectionsystem,while\NP"mark edslicesconsideredtocon tainpathology ,butno enhancingtumorb ythepost-processingstep.Inallv olumes:TP=T rueP ositiv e;FP =F alseP ositiv e;FN=F alseNegativ e;%M=P ercen tMatc h;CR=Correspondence Ratio. PATIENT1 BASELINE Sl.Tr. TP FP FN %M CR 18 56 202 0 1.00 -0.80 #Slices 10 19C 696 10 7 0.99 0.98 KBVol: 7970 20 1048 31 13 0.99 0.97 GTVol: 7213 21 1089 34 27 0.98 0.96 22 857 20 22 0.97 0.96 23 771 113 9 0.99 0.92 24 838 61 23 0.97 0.94 25 705 246 1 1.00 0.82 26 610 101 31 0.95 0.87 27 288 194 122 0.70 0.47 TOTAL: 69581012 255 0.96 0.89

PAGE 205

APPENDIXB(Con tin ued) 190 PATIENT1 REPEAT1 Sl. Tr. TP FP FN %M CR 16 6 38 0 1.00 -2.17 #Slices 11 17C 428 95 8 0.98 0.87 KBVol: 8893 18 995 0 55 0.95 0.95 GTVol: 7240 19 1098 11 62 0.95 0.94 20 902 95 13 0.99 0.93 21 TS5 694 32 41 0.94 0.92 22 789 102 14 0.98 0.92 23 764 221 5 0.99 0.85 24 664 109 29 0.96 0.88 25 409 391 21 0.95 0.50 26 191 859 52 0.79 -0.98 TOTAL: 69401953 300 0.96 0.82 REPEAT2 Sl. Tr. TP FP FN %M CR 18 1 228 2 0.33-37.67 #Slices 11 19C 352 90 1 1.00 0.87 KBVol: 9281 20 979 71 0 1.00 0.96 GTVol: 7470 21 1233 209 5 1.00 0.91 22 1001 95 5 1.00 0.95 23 764 192 16 0.98 0.86 24 808 130 4 1.00 0.92 25 751 15 76 0.91 0.90 26 723 193 3 1.00 0.86 27 453 288 1 1.00 0.68 28 150 555 142 0.51 -0.44 TOTAL: 72151838 255 0.97 0.83

PAGE 206

APPENDIXB(Con tin ued) 191 PATIENT1 REPEAT3 Sl. Tr. TP FP FN %M CR 17 20 0 3 0.87 0.87 #Slices 10 18C 528 134 0 1.00 0.87 KBVol: 7539 19 1017 31 8 0.99 0.98 GTVol: 6395 20 985 66 17 0.98 0.95 21 725 131 8 0.99 0.90 22 683 30 71 0.91 0.89 23 734 154 12 0.98 0.88 24 770 168 1 1.00 0.89 25 430 391 0 1.00 0.55 26 329 213 54 0.86 0.58 TOTAL: 62211318 174 0.97 0.87 REPEAT4 Sl. Tr. TP FP FN %M CR 15 0 355 8 0.00-22.19 #Slices 12 16 7 177 0 1.00-11.64 KBVol: 8314 17 208 174 1 1.00 0.58 GTVol: 6560 18C 732 59 8 0.99 0.95 19 1020 109 13 0.99 0.93 20 922 181 16 0.98 0.89 21 721 81 18 0.98 0.92 22 699 49 3 1.00 0.96 23 733 239 2 1.00 0.83 24 689 190 5 0.99 0.86 25 396 271 2 0.99 0.65 26 199 103 158 0.56 0.41 TOTAL: 63261988 234 0.96 0.81

PAGE 207

APPENDIXB(Con tin ued) 192 PATIENT2 BASELINE Sl.Tr. TP FP FN %M CR 11 TS5 522 339 41 0.93 0.63 #Slices 13 12 TS5 909 444 65 0.93 0.71 KBVol:16214 13 TS5 1292 452 85 0.94 0.77 GTVol:12230 14 TS5 770 540 22 0.97 0.63 15CTS5 909 527 32 0.97 0.69 16 TS5 1354 308 43 0.97 0.86 17 TS5 1374 244 271 0.84 0.76 18 TS5 1217 727 74 0.94 0.66 19 TS5 946 407 95 0.91 0.71 20 TS5 850 544 55 0.94 0.64 21 TS5 474 122 65 0.88 0.77 22 TS5 460 203 81 0.85 0.66 23 TS5 200 80 24 0.89 0.71 TOTAL: 112774937 953 0.92 0.72 REPEAT1 Sl. Tr. TP FP FN %M CR 11 All 177 24 49 0.78 0.73 #Slices 14 12 All 517 128 86 0.86 0.75 KBVol:15623 13 All 763 118 110 0.87 0.81 GTVol:14609 14 All 1319 112 273 0.83 0.79 15C AT 1198 138 193 0.86 0.81 16 AT 1246 249 155 0.89 0.80 17 AT 1577 185 224 0.88 0.82 18 AT 1913 235 199 0.91 0.85 19 AT 1650 728 102 0.94 0.73 20 AT 1259 494 166 0.88 0.71 21 AT 505 217 61 0.89 0.70 22 AT 292 292 55 0.84 0.42 23 AT 272 15 83 0.77 0.75 24 AT 0 0 165 0.00 0.00NP TOTAL: 1268829351756 0.87 0.77

PAGE 208

APPENDIXB(Con tin ued) 193 PATIENT2 REPEAT2 Sl. Tr. TP FP FN %M CR 11 328 23 47 0.87 0.84 #Slices 15 12 300 336 308 0.49 0.22 KBVol:22255 13 TS5 1517 180 196 0.89 0.83 GTVol:20924 14 1783 305 156 0.92 0.84 15C 1295 114 371 0.78 0.74 16 1936 419 86 0.96 0.85 17 TS5 2085 277 336 0.86 0.80 18 2814 454 372 0.88 0.81 19 2338 614 146 0.94 0.82 20 1614 670 94 0.94 0.75 21 1223 392 118 0.91 0.77 22 556 154 276 0.67 0.58 23 398 130 47 0.89 0.75 24 0 0 172 0.00 0.00NP 25 0 0 12 0.00 0.00 N TOTAL: 1818740682737 0.87 0.77

PAGE 209

APPENDIXB(Con tin ued) 194 PATIENT3 BASELINE Sl.Tr. TP FP FN %M CR 13 105 348 0 1.00 -0.66 #Slices 12 14 595 74 22 0.96 0.90 KBVol:10854 15 664 221 8 0.99 0.82 GTVol:10892 16C 863 79 32 0.96 0.92 17 891 49 54 0.94 0.92 18 1135 3 121 0.90 0.90 19 1167 0 195 0.86 0.86 20 1341 8 179 0.88 0.88 21 1116 78 38 0.97 0.93 22 949 2 139 0.87 0.87 23 696 9 178 0.80 0.79 24 404 57 0 1.00 0.93 TOTAL: 9926 928 966 0.91 0.87 REPEAT1 Sl. Tr. TP FP FN %M CR 13 0 0 13 0.00 0.00 N #Slices 12 14 252 277 5 0.98 0.44 KBVol: 7387 15 511 216 14 0.97 0.77 GTVol: 5971 16C 731 121 33 0.96 0.88 17 722 8 88 0.89 0.89 18 831 249 10 0.99 0.84 19 821 337 9 0.99 0.79 20 736 178 9 0.99 0.87 21 495 30 19 0.96 0.93 22 399 97 25 0.94 0.83 23 176 89 9 0.95 0.71 24 63 48 0 1.00 0.62 25 0 0 0 NP TOTAL: 57371650 234 0.96 0.82

PAGE 210

APPENDIXB(Con tin ued) 195 PATIENT4 BASELINE Sl.Tr. TP FP FN %M CR 09L P 179 8 194 0.48 0.47 #Slices 16 10 377 81 78 0.83 0.74 KBVol:11827 11 839 181 140 0.86 0.76 GTVol:10454 12 867 253 97 0.90 0.77 13 961 180 45 0.96 0.87 14 832 896 153 0.84 0.39 15 569 26 140 0.80 0.78 16 709 22 116 0.86 0.85 17 832 86 162 0.84 0.79 18C 726 154 23 0.97 0.87 19 732 350 12 0.98 0.75 20 813 74 95 0.90 0.85 21 368 77 7 0.98 0.88 22 271 113 3 0.99 0.78 23 79 44 0 1.00 0.72 24 30 98 5 0.86 -0.54 TOTAL: 915425451265 0.88 0.75 REPEAT1 Sl. Tr. TP FP FN %M CR 11L P 147 32 16 0.90 0.80 #Slices 15 12 268 48 125 0.68 0.62 KBVol:10007 13 631 96 227 0.74 0.68 GTVol:10835 14 713 35 304 0.70 0.68 15 1280 87 152 0.89 0.86 16 928 27 92 0.91 0.90 17 669 185 88 0.88 0.76 18 739 12 157 0.82 0.82 19 807 14 192 0.81 0.80 20C 785 77 203 0.79 0.76 21 TS5 762 40 197 0.79 0.77 22 541 47 327 0.62 0.60 23 198 284 1 0.99 0.28 24 204 93 34 0.86 0.66 25 0 258 48 0.00 -2.69 TOTAL: 867213352163 0.80 0.74

PAGE 211

APPENDIXB(Con tin ued) 196 PATIENT4 REPEAT2 Sl. Tr. TP FP FN %M CR 10L P 213 142 68 0.76 0.51 #Slices 15 11 P 671 83 199 0.77 0.72 KBVol:16124 12 All 1580 278 106 0.94 0.85 GTVol:15788 13 All 1848 182 197 0.90 0.86 14 All 2182 186 250 0.90 0.86 15 All 1618 213 195 0.89 0.83 16 All 1800 279 119 0.94 0.87 17 All 1063 277 79 0.93 0.81 18 All 925 143 116 0.89 0.82 19C AT 589 79 285 0.67 0.63 20 AT 688 37 132 0.84 0.82 21 AT 282 151 87 0.76 0.56 22 AT 191 98 76 0.72 0.53 23 AT 126 123 55 0.70 0.36 24 AT 48 29 0 1.00 0.70 TOTAL: 1382423001964 0.88 0.80

PAGE 212

APPENDIXB(Con tin ued) 197 PATIENT5 BASELINE Sl. Tr. TP FP FN %M CR 07LAll 422 170 16 0.96 0.77 #Slices 9 08LAll 967 130 72 0.93 0.87 KBVol:10739 09 All 1623 303 49 0.97 0.88 GTVol:10178 10 All 1735 151 119 0.94 0.90 11 All 1704 67 102 0.94 0.92 12 All 1542 62 111 0.93 0.91 13 All 833 70 10 0.99 0.95 14 All 674 38 22 0.97 0.94 15CAll 175 73 2 0.99 0.78 TOTAL: 96751064 503 0.95 0.90 REPEAT1 Sl. Tr. TP FP FN %M CR 07L P 0 0 0 N #Slices 10 08L P 60 130 6 0.91 -0.08 KBVol: 5935 09 899 553 67 0.93 0.64 GTVol: 4657 10 931 311 6 0.99 0.83 11 898 100 13 0.99 0.93 12 913 112 18 0.98 0.92 13 702 108 9 0.99 0.91 14 134 84 1 0.99 0.68 15C 0 0 0 N 16 0 0 0 N TOTAL: 45371398 120 0.97 0.82 REPEAT2 Sl. Tr. TP FP FN %M CR 07L 57 243 1 0.98 -1.11 #Slices 8 08LTS5 529 202 187 0.74 0.60 KBVol: 6302 09 987 283 122 0.89 0.76 GTVol: 5616 10 874 12 285 0.75 0.75 11 1302 13 27 0.98 0.97 12 979 329 1 1.00 0.83 13 265 227 0 1.00 0.57 14 0 0 0 N TOTAL: 49931309 623 0.89 0.77

PAGE 213

APPENDIXB(Con tin ued) 198 PATIENT5 REPEAT3 Sl. Tr. TP FP FN %M CR 07L P 91 369 28 0.76 -0.79 #Slices 8 08L 629 145 123 0.84 0.74 KBVol: 9612 09 AT 1175 162 92 0.93 0.86 GTVol: 9215 10 1451 93 24 0.98 0.95 11 1366 10 185 0.88 0.88 12 1664 31 154 0.92 0.91 13 1276 15 47 0.96 0.96 14 872 89 38 0.96 0.91 15 0 174 0 16C 0 0 0 N TOTAL: 85241088 691 0.93 0.88 REPEAT4 Sl. Tr. TP FP FN %M CR 08LTS5 248 271 153 0.62 0.28 #Slices 7 09L 136 172 20 0.87 0.32 KBVol: 4957 10 358 143 47 0.88 0.71 GTVol: 3544 11 1051 387 65 0.94 0.77 12 788 346 4 0.99 0.78 13 539 131 53 0.91 0.80 14 78 309 4 0.95 -0.93 15C 0 0 0 TOTAL: 31981759 346 0.90 0.65

PAGE 214

APPENDIXB(Con tin ued) 199 PATIENT6 BASELINE Sl. Tr. TP FP FN %M CR 13 169 360 4 0.98 -0.06 #Slices 15 14 370 458 10 0.97 0.37 KBVol:11266 15 435 327 14 0.97 0.60 GTVol: 5853 16 362 871 29 0.93 -0.19 17 571 872 5 0.99 0.23 18C 570 183 67 0.83 0.71 19 495 621 13 0.94 0.28 20 615 365 0 1.00 0.70 21 296 388 0 1.00 0.40 22 554 433 1 0.99 0.55 23 510 542 1 0.99 0.47 24 427 296 25 0.94 0.62 25 150 26 64 0.70 0.64 26 0 0 72 0.00 0.00NP 27 0 0 0 NP TOTAL: 54825784 371 0.94 0.44 REPEAT1 Sl. Tr. TP FP FN %M CR 11 53 338 0 1.00 -2.19 #Slices 12 12 76 715 3 0.96 -3.56 KBVol: 9105 13 188 414 1 0.99 -0.10 GTVol: 2754 14 360 757 2 0.99 -0.05 15 207 940 4 0.98 -1.25 16C 658 883 1 1.00 0.33 17 250 478 7 0.98 0.04 18 404 656 1 1.00 0.19 19 88 401 9 0.92 -1.16 20 92 438 0 1.00 -1.38 21 97 108 7 0.93 0.41 22 161 343 15 0.91 -0.06 TOTAL: 27046401 50 0.98 -0.18

PAGE 215

APPENDIXB(Con tin ued) 200 PATIENT6 REPEAT2 Sl. Tr. TP FP FN %M CR 12 0 0 0 N 13 69 928 272 0.20 -1.16 #Slices 13 14 0 901 39 0.00-11.55 KBVol: 9821 15 402 487 14 0.97 0.38 GTVol: 4354 16 525 318 22 0.96 0.67 17 369 868 3 0.99 -0.17 18C 612 77 102 0.86 0.80 19 641 802 7 0.99 0.37 20 384 83 67 0.85 0.76 21 351 558 0 1.00 0.21 22 311 249 0 1.00 0.60 23 81 487 1 0.99 -1.98 24 76 242 6 0.93 -0.55 TOTAL: 38216000 533 0.88 0.19 REPEAT3 Sl. Tr. TP FP FN %M CR 11 0 295 0 #Slices 15 12 224 603 0 1.00 -0.35 KBVol:12234 13 376 837 0 1.00 -0.11 GTVol: 6510 14 572 731 1 1.00 0.36 15 637 897 0 1.00 0.30 16 839 844 1 1.00 0.50 17C 733 58 139 0.84 0.81 18 1015 456 7 0.99 0.77 19 392 146 76 0.84 0.68 20 561 343 0 1.00 0.69 21 354 344 3 0.99 0.51 22 214 394 0 1.00 0.08 23 166 73 9 0.95 0.74 24 60 53 111 0.35 0.20 25 0 17 20 0.00 -0.42 TOTAL: 61436091 367 0.94 0.50

PAGE 216

APPENDIXB(Con tin ued) 201 PATIENT6 REPEAT4 Sl. Tr. TP FP FN %M CR 10 0 650 0 #Slices 15 11 154 408 1 0.99 -0.32 KBVol:12444 12 900 429 23 0.98 0.74 GTVol: 8688 13 847 837 51 0.94 0.48 14 954 102 172 0.85 0.80 15 762 944 38 0.95 0.36 16C 1014 146 78 0.93 0.86 17 775 162 80 0.91 0.81 18 786 28 137 0.85 0.84 19 502 37 176 0.74 0.71 20 446 209 140 0.76 0.58 21 436 227 22 0.95 0.70 22 192 464 2 0.99 -0.21 23 0 33 0 24 0 0 0 TOTAL: 77684676 920 0.89 0.66 REPEAT5 Sl. Tr. TP FP FN %M CR 10 0 0 0 N 11 0 0 0 N 12 8 84 0 1.00 -4.25 #Slices 14 13 0 0 41 0.00 0.00 N KBVol: 5507 14 01251 205 0.00 -3.05 GTVol: 3079 15 273 423 7 0.97 0.22 16 379 324 9 0.98 0.56 17C 370 81 64 0.85 0.76 18 334 1 258 0.56 0.56 19 430 177 23 0.95 0.75 20 295 223 11 0.96 0.60 21 194 181 20 0.91 0.48 22 124 143 2 0.98 0.42 23 25 119 7 0.78 -1.08 24 0 68 0 25 0 0 0 NP TOTAL: 24323075 647 0.79 0.30

PAGE 217

APPENDIXB(Con tin ued) 202 PATIENT6 REPEAT6 Sl. Tr. TP FP FN %M CR 11 0 0 8 0.00 0.00 N #Slices 15 12 0 0 13 0.00 0.00 N KBVol: 6462 13 55 152 0 1.00 -0.38 GTVol: 4384 14 282 300 5 0.98 0.46 15 479 544 4 0.99 0.43 16C 719 562 3 1.00 0.61 17 616 34 138 0.82 0.79 18 562 30 118 0.83 0.80 19 487 47 148 0.77 0.73 20 412 220 3 0.99 0.73 21 320 237 4 0.99 0.62 22 8 225 0 1.00-13.06 23 0 94 0 24 0 77 0 25 0 0 0 NP TOTAL: 39402522 444 0.90 0.63 REPEAT7 Sl. Tr. TP FP FN %M CR 10 0 65 0 #Slices 14 11 0 42 0 KBVol: 5647 12 75 10 25 0.75 0.70 GTVol: 4047 13 146 37 29 0.83 0.73 14 252 327 9 0.97 0.34 15 461 464 4 0.99 0.49 16 662 586 60 0.92 0.51 17C 503 36 138 0.78 0.76 18 521 158 24 0.96 0.81 19 375 120 115 0.77 0.64 20 338 55 93 0.78 0.72 21 177 218 4 0.98 0.38 22 13 6 23 0.36 0.28 23 0 0 0 TOTAL: 35232124 524 0.87 0.62

PAGE 218

APPENDIXB(Con tin ued) 203 PATIENT6 REPEAT8 Sl. Tr. TP FP FN %M CR 10 26 91 5 0.84 -0.63 #Slices 14 11 56 26 9 0.86 0.66 KBVol: 4849 12 145 4 29 0.83 0.82 GTVol: 3935 13 243 235 19 0.93 0.48 14 404 578 16 0.96 0.27 15 549 123 159 0.78 0.69 16C 512 46 155 0.77 0.73 17 400 29 106 0.79 0.76 18 304 63 61 0.83 0.75 19 308 64 86 0.78 0.70 20 188 54 58 0.76 0.65 21 35 221 45 0.44 -0.94 22 10 135 7 0.59 -3.38 23 0 0 0 NP TOTAL: 31801669 755 0.81 0.60 REPEAT9 Sl. Tr. TP FP FN %M CR 11 44 246 1 0.98 -1.76 #Slices 11 12 117 85 29 0.80 0.51 KBVol: 4932 13 114 143 50 0.70 0.26 GTVol: 3201 14 399 558 24 0.94 0.28 15C 439 524 155 0.74 0.30 16 460 39 136 0.77 0.74 17 343 39 84 0.80 0.76 18 331 231 22 0.94 0.61 19 231 73 41 0.85 0.72 20 152 151 7 0.96 0.48 21 22 191 0 1.00 -3.34 TOTAL: 26522280 549 0.83 0.47

PAGE 219

APPENDIXB(Con tin ued) 204 PATIENT6 REPEAT10 Sl. Tr. TP FP FN %M CR 10 136 96 32 0.81 0.52 #Slices 14 11 156 11 32 0.83 0.80 KBVol: 5081 12 207 61 100 0.67 0.57 GTVol: 3895 13 241 356 3 0.99 0.26 14 303 279 1 1.00 0.54 15 301 479 14 0.96 0.20 16 678 52 133 0.84 0.80 17 350 49 73 0.83 0.77 18C 292 94 104 0.74 0.62 19 211 49 57 0.79 0.70 20 276 122 11 0.96 0.75 21 114 57 27 0.81 0.61 22 35 76 8 0.81 -0.07 23 0 0 0 NP TOTAL: 33001781 595 0.85 0.62 REPEAT11 Sl. Tr. TP FP FN %M CR 07L 0 0 8 0.00 0.00 N #Slices 18 08L 0 0 83 0.00 0.00 N KBVol: 8246 09 223 123 23 0.91 0.66 GTVol: 5691 10 568 208 8 0.99 0.81 11 520 344 7 0.99 0.66 12 593 469 9 0.99 0.60 13 423 55 61 0.87 0.82 14 453 337 5 0.99 0.62 15 561 501 22 0.96 0.53 16 665 51 111 0.86 0.82 17C 397 287 18 0.96 0.61 18 284 16 65 0.81 0.79 19 183 96 60 0.75 0.56 20 218 169 21 0.91 0.56 21 30 256 0 1.00 -3.27 22 32 178 27 0.54 -0.97 23 6 0 7 0.46 0.46 24 0 0 0 NP TOTAL: 51563090 535 0.91 0.63

PAGE 220

APPENDIXB(Con tin ued) 205 PATIENT6 REPEAT12 Sl. Tr. TP FP FN %M CR 07L 19 47 47 0.29 -0.07 #Slices 18 08L 203 255 78 0.72 0.27 KBVol: 9274 09 201 271 67 0.75 0.24 GTVol: 6438 10 420 692 112 0.79 0.14 11 597 155 116 0.84 0.73 12 653 202 23 0.97 0.82 13 406 393 80 0.84 0.43 14 292 327 33 0.90 0.40 15 346 580 126 0.73 0.12 16 878 169 97 0.90 0.81 17C 409 261 51 0.89 0.61 18 337 79 125 0.73 0.64 19 202 298 18 0.92 0.24 20 186 23 48 0.79 0.75 21 126 161 25 0.83 0.30 22 56 30 31 0.64 0.47 23 0 0 26 0.00 0.00NP 24 0 0 4 0.00 0.00NP TOTAL: 533139431107 0.83 0.52

PAGE 221

APPENDIXB(Con tin ued) 206 PATIENT7 BASELINE Sl. Tr. TP FP FN %M CR 15 0 0 0 N #Slices 12 16 0 0 0 N KBVol: 1200 17 0 0 0 N GTVol: 1018 18 ALL 39 58 7 0.85 0.22 19C ALL 106 68 10 0.91 0.62 20 ALL 285 33 30 0.90 0.85 21 ALL 210 20 60 0.78 0.74 22 ALL 162 100 23 0.88 0.61 23 ALL 25 12 45 0.36 0.27 24 ALL 15 67 1 0.94 -1.16 25 0 0 0 NP 26 0 0 0 NP TOTAL: 842 358 176 0.83 0.65 PATIENT8 BASELINE Sl.Tr. TP FP FN %M CR 14 195 248 1 0.99 0.36 #Slices 7 15 757 150 23 0.97 0.87 KBVol: 4839 16 987 11 74 0.93 0.93 GTVol: 4521 17 983 8 76 0.93 0.92 18C 854 15 20 0.98 0.97 19 472 28 18 0.96 0.93 20 61 70 0 1.00 0.43 21 0 0 0 NP 22 0 289 0 23 0 0 0 NP 24 0 0 0 NP TOTAL: 4309 530 212 0.95 0.89

PAGE 222

207 APPENDIXC EFF OR TSUSINGEDGEDETECTION Asmen tionedinChapter7,oneofthestrengthsofthekno wledge-basedapproac hisitsabilit ytoev olv easnewdomaininformationbecomesa v ailableandprocessingtoolsaredev eloped.Oneoftheseinthepreliminarystagesofdev elopmen tis theuseofedgedetectiontoaidthesettingofthenalthresholdinStageFiv e.Lik e mostoftheedgedetectionmethodsdescribedinChapter3,thetec hniqueisbasedon theprinciplethatedgescanbefoundalongtheborderbet w eenenhancingtumorand thesurroundingtissues.Themoredistincttheboundaryis,thegreaterthein tensit y dierence(usuallyinT1-w eigh tedspace,wheregadoliniumenhancemen tismosteffectiv e)andthestrongeranedgeitwillha v e.Mostedge-detectionbasedmethods, suc hasthosedescribedinSection3.1.3,attempttouseedgestrengthtracethetumor'scon tours.Thiscanw orkw ellfortumorswithdistinctboundaries,assho wnin Figure47(d)usingastandardSobeloperator,butcanha v esignican tproblemswith morediusetumors,sho wninFigure47(i). An um berofedgedetectionoperatorsha v ebeenin troduced,suc hasCann yand Bergholmmostoftheseha v ean um berofparametersthatarediculttoautomatically optimize,especiallyinadomainwheretheobjectofin terestcanha v esuc hwide rangingc haracteristics.Asaresult,Dellepiane[30]andRaandNewman[96]ha v e suggestedthatedgedetectionareunlik elytow orkreliablyforcomplexstructureslik e tumors.Edgedetectionma ystillpro videkno wledgethattobeexploited,ho w ev er. Bynotingthatedgesnotonlyappro ximatethetumorboundaryspatially ,butcan alsoindicatetheappro ximatesignalin tensit yofthatboundary ,whic hcanbeusedin athresholdoperation.Thisconceptismoreexiblesincedetectededgeneednotbe perfect,merelysucien ttoindicatetheappropriatesignalin tensit y .Edgedetection m uststillbereasonable,ho w ev er,forthemethodtow ork.

PAGE 223

APPENDIXB(Con tin ued) 208 T oaddresstheproblemofdetectingedgesintumorswithdiuse,andtominimizetheproblemofparameteroptimization,thetec hniquein troducedhereusesa \fuzzy"approac htoedgedetectionpresen tedb yT aoandThompsonin[109]whic h usedfuzzyif-thenrulesthatw erebasedontherelationshipbet w eeneac hpixeland itseigh t-wiseneigh bors.Structureelemen ts,sixteenintotal,examplessho wninFigure48,areusedtodev elopafuzzyif-thenrule: IF [thedierences( D x 's)bet w eenthein tensitiesofthepixels(mark edwith\x")and thecen terpixelaresmall] AND [thedierences( D 's)bet w eenthein tensitiesofthe pixels(notmark edwith\x")andthecen terpixelarelarge] THEN thecen terpixel ofthisstructureisanedgepixel. Theauthorsstatethatthefuzzymem berships small and large aredenedwith abell-shapedfunction,thoughtheydonotspecifyaparticularfunction. Here,a Gaussian-basedfunctionisusedandthefuzzyset small isdenedas: small = e )Tj/T10 1 Tf32.0001 -32 TD1 Tc[(Diff 2 ( a;b ) 2 2 where Diff ( a;b ) istheabsolutein tensit ydierencebet w eenthecen terpixelandthe eigh t-wiseneigh bor(a,b).Thefuzzyset large isdenedas: large =1 )Tj/T17 1 Tf49.9999 0 TD( small .The actualGaussianform ulaisnotusedtoallo wamem bershipof small =1 : 0tobe returnedwhen Diff =0.Also,inastandardGaussianfunction,thev alue represen tsthestandarddeviation.F ordeningthefuzzyset,itcon trolsho wquic kly small decreases(and large increases)asthein tensit ydierence, Diff ,becomeslarger.In thispreliminarystudy =2 : 0,thoughitcouldbepossibletoha v erulesinthe kno wledge-baseadjustthev alueaccordingtoatumor'sc haracteristics. Thefuzzyif-thenruledescribedabo v eisusedtodetermineapixel's\edge poten tial"(PEP)withinagiv enedgestructure(Figure48)b ycalculatingthemembershipsbet w eenthecen terpixelandeac hofiteigh t-wiseneigh borsforthefuzzy

PAGE 224

APPENDIXB(Con tin ued) 209 (a)Ra wData (b)GTT umor (c)KBT umor (d)SobelEdge (e)F uzzyEdge (f)Ra wData (g)GTT umor (h)KBT umor (i)SobelEdge (j)F uzzyEdge Figure47.DetectingEdgesAlongT umorBoundaries.Giv enatumorsegmen tation mask,(b)and(h),producedb yStageF our,edgedetectionisperformed onpixelscon tainedb ythemasktondthetumor'sboundaries. The resultsofaSobeloperatoraresho wnin(d)and(i),while(e)and(j) sho wtheresultsofafuzzyedgedetectordescribedin[109].Thetumor in(a)hasdistinctedgesandbothedgedetectorsw orkw ell,thoughthe Sobeloperatormorecloselymatc hesgroundtruth(b).Thetumorin(e), ho w ev er,performsrelativ elypoorly ,forboththeSobelandfuzzyedge detectors.

PAGE 225

APPENDIXB(Con tin ued) 210 X X X X X X X X XX X X X X X X X X X X X X XXFigure48.EdgeStructuresforF uzzyEdgeDetection. Edgestructures,examples sho wnabo v e,areusedtogeneratethefuzzyif-thenrules. Neigh boring pixelsco v eredb yan\x"calculatemem bershipinthefuzzyset small whilethoseunco v eredcalculatemem bershipinthefuzzyset large set small (iftheneigh borisco v eredb yan\x"intheedgestructure)or large (ifthe neigh borisunco v ered),andreturningtheminim ummem bership.F orexample,using therststructure(PEP1)inFigure48,itsedgepoten tialw ouldbe: PEP 1 ( x;y )=min ( small ( Diff ( x )Tj/T15 1 Tf27 0 TD(1 ;y )Tj/T15 1 Tf27.0001 0 TD(1) ) ; small ( Diff ( x )Tj/T15 1 Tf27 0 TD(1 ;y )Tj/T15 1 Tf27.0001 0 TD(1) ) ; small ( Diff ( x )Tj/T15 1 Tf26.9998 0 TD(1 ;y ) ) ; small ( Diff ( x )Tj/T15 1 Tf27.0001 0 TD(1 ;y +1) ) ; small ( Diff ( x;y )Tj/T15 1 Tf27 0 TD(1) ) ; small ( Diff ( x;y +1) ) ; large ( Diff ( x +1 ;y )Tj/T15 1 Tf26.9999 0 TD(1) ) ; large ( Diff ( x +1 ;y ) ) ; large ( Diff ( x +1 ;y +1) )) Giv ensixteenedgestructures,apixelwillha v esixteenpossibleedgememberships. Thepixel'snaledgemem bershipissetb yk eepingmem bershipofthe structurethebestmatc hedtheedge(i.e.,thestructurewiththehighestmem bership).F ormally: PEP ( x;y )=max ( PEP 1 ( x;y ) ;:::; PEP 16 ( x;y )) Oncenaledgemem bershipsha v ebeencalculatedforallpixels,thedetected edgesare\thinned"b yremo vingredundan tedgepixelsthroughalocalmaxima operation.A\pseudo-cen troid"oftheremainingedgestrengthsisthencalculatedand onlythoseedgesthatarestrongerthanthepseudo-cen troidarek ept.Themethod proposedb yT aoandThompsonw asimplem e n tedasdescribedin[109]withthe additionthatthetec hniqueonlyconsiderspixelscon tainedinanimagemask(inthis case,thetumorsegmen tationmaskproducedattheendofStageF our).

PAGE 226

APPENDIXB(Con tin ued) 211 T able18.FinalThresholding: \StageFiv e"Vs. F uzzy-EdgeBased. T umorv olumesgeneratedb ytherulesdescribedinSection5.5andfuzzy-edgebased thresholdingarecompared. P atien t Scan Threshold T rue F alse F alse P ercen t Corr. Method P ositiv e P ositiv e Negativ e Matc h Ratio 1 R1 Rule 6940 1953 300 0.96 0.82 Edge 7071 1758 169 0.98 0.86 2 R1 Rule 12688 2935 1756 0.87 0.77 Edge 11096 1778 3514 0.76 0.70 2 R2 Rule 18187 4068 2737 0.87 0.77 Edge 16676 2865 4249 0.80 0.73 ResultsofthefuzzyedgedetectionmethodcanbeseeninFigures47(e)and(j) fortumorswithdistinctanddiuseboundariesrespectiv ely .Asmen tionedearlier, ho w ev er,themethodbeinginitiallyin v estigateddoesnotrequiretheexacttumor boundaries,onlytheirappro ximation,sothatanalthresholdma ybesetbasedon thein tensitiesofthepixelscon tainedinthoseedges.Giv enafuzzyedgeimage,a thresholdissetb ycalculatingthemeanin tensit y(inT1-w eigh tedspace)ofeac hpixel con tainedinthefuzzyedgeimage.Now eigh ting(accordingtoedgemem bership)of in tensitiesisperformedasexperimen tssho w edthatw eigh tingdidnotsignican tly aecttheresultan tthresholdv alue. T able18sho wssomeexampletumorv olumes usingthisfuzzy-edgebasedthresholding.Ascanbeseen,whileperformingw ellin caseswithdistincttumorboundaries,themethodhasproblemswithmorediuse cases,suc hasP atien t2.Themethodhaspromise,ho w ev er,andshouldbefurther in v estigated.

PAGE 227

212 APPENDIXD R ULESINCLIPS DuetothesizeoftheCLIPSrulesandmodulesforimageprocessingand m ultispectralanalysisusedinthisw ork,theycouldnotbeincludedinhardcop y format.Instead,theyha v ebeenincludedonCD-R OMinIS0-9660format(readable onUnix,DOS,andMacplatforms).

PAGE 228

VIT A MatthewCaseyClarkreceiv edhisMaster'sdegreein1994inComputerScience andEngineeringfromtheUniv ersit yofSouthFloridainT ampa,Florida.Thesubject ofthisMaster'sThesisw astheprocessingofpartialmagneticresonancev olumesofthe h umanbraintodetectpathologyandlabelnormaltissues.Sincethen,hehasw ork ed inthedepartmen tofComputerScienceandEngineeringandwiththedepartmen t ofRadiologyindev elopmen tofasystemtosegmen tbraintumorsinMRimages. Hehasv epublications,includinganarticlein IEEE:EngineeringinMedicineand Biology .Hiscurren tresearc hin terestsarein telligen tsystemsandtheuseofarticial in telligencetec hniquesinpatternrecognitionandimageprocessingapplications.