Support-Vector-Machine-Approach-for-Classification

Abstract—Theobjectiveofthispaper,istoapplysupportvectormachine(SVM)approachfortheclassificationofcancerousandnormalregionsofprostateimages.Threekindsoftexturalfeaturesareextractedandusedfortheanalysis:parametersoftheGauss-Markovrandomfield(GMRF),correlationfunctionandrelativeentropy.Prostateimagesareacquiredbythesystemconsistingofamicroscope,videocameraandadigitizingboard.Cross-validatedclassificationoveradatabaseof46imagesisimplementedtoevaluatetheperformance.InSVMclassification,sensitivityandspecificityof96.2%and97.0%areachievedforthe32x32pixelblocksizeddata,respectively,withanoverallaccuracyof96.6%.Classificationperformanceiscomparedwithartificialneuralnetworkandk-nearestneighborclassifiers.ExperimentalresultsdemonstratethattheSVMapproachgivesthebestperformance.Keywords—Computer-aideddiagnosis,supportvectormachines,Gauss-Markovrandomfields,textureclassification.I.INTRODUCTIONARLYdetectionofcancerisveryimportantforsavinglives.Prostatecancerisoneofthecommonlydiagnosedcancerinmen.Diagnosisofprostatecancerrequiresthetissueandcellspecimens.Thesespecimens(asgiveninFig.1)arescreenedandanalyzedbyapathologistusingamicroscope.Optimummedicaltreatmentisdecidedaccordingtothisinformationreportedbythepathologist.Wrongdiagnosisofcancercasesisamajorprobleminpathology.Insomecases,correctdiagnosisisveryhardandtherecanbe30-40%differencebetweenpathologists’decisions[1].Examplesofdiagnosticerrorsofbiopsyslidesaregivenin[2].Forthesolutionofthisproblem,computeraideddiagnosis(CAD)systemscanbeused.ACADsystem’sgoalistoincreasethepathologists’performanceandaccuracybyindicatingtheregionsofabnormalities(e.g.cancerousregions).Quantitativeanalysisresultsofprostateimagescanreducethenumberofmisclassificationsandimprovethediagnosisperformance.Inthiswork,supportvectormachine(SVM)learningandclassificationisinvestigatedforthedetectionofcancerousprostateregions.ItisshownthatSVMprovidesgoodperformanceamongthetestedmethods.SVMtrainingisThisworkisfundedbyDokuzEylülUniv.,BAPundergrant02.KB.FEN.058.M.MakinacıiswiththeElectricalandElectronicsEngineeringDepartment,DokuzEylülUniversity,İzmir,Turkey.(phone:[phone];fax:[phone];e-mail:[e-mail]).basedonthestatisticallearningtheory[3].Inrecentyears,SVMlearningiswidelyusedinbiomedicalapplications,includingmicrocalcificationdetection[4],rheumatoidjointinflammationclassification[5],fetallungmaturityanalysis[6],andpolypdetection[7].TheuseofSVMinclassificationtaskshassomeadvantages.Comparedwiththeneuralnetwork(NN)approaches,findingageneralNNclassifiermodeltofitanydataisverydifficult.ThisdifficultyisnotaconcernwhentheSVMclassifierisused.Inthisstudy,classificationofcancerousandnormalprostateregionsisdemonstratedasatwoclasspatternrecognitiontask.SupervisedlearningisusedtotrainthenonlinearSVMclassifier.InSectionIIimageacquisitionsystemandhistologyofthespecimensarepresented.FeatureextractionandclassificationproceduresaregiveninSectionIIIandIV,respectively.InSectionV,performanceanalysismethodsaresummarized,andclassificationresultsareevaluatedinSectionVI.SomeconclusionsandfurtherresearchdirectionsareprovidedinSectionVII.II.MATERIALSA.ImageAcquisitionThespecimenimagesarex100magnifiedbytheLeicamicroscope.Anoilimmersionobjectiveisused.Theanalogimagesignalisacquiredwithacolorcameraandwithas-videoconnectionthesignalistransmittedtothecomputer.Theimagesareacquiredat480x360pixel24bit/pixelformatandsaved.B.HistologyMicroscopicslicesaretheradicalprostatectomymaterialsbelongingtothepatientswhohadasurgicaloperationbecauseofprostatecancer.Theradicalprostatectomymaterialsarefixedin10%formalinfor24-48h.topreservethebiologicstructure.Routinetissueinspectionisappliedtotheentirestructure.Paraffinblockspreparedfromthetissuecutinto5-µmthickslicesandstainedwithheamatoxylinandeosin.MetehanMakinacıSupportVectorMachineApproachforClassificationofCancerousProstateRegionsEPROCEEDINGSOFWORLDACADEMYOFSCIENCE,ENGINEERINGANDTECHNOLOGYVOLUME7AUGUST2005ISSN1307-6884PWASETVOLUME7AUGUST2005ISSN1307-6884166©2005WASET.ORGFig.1Prostateimage.Attherightnormalprostaticglandsandattheleftprostaticadenocarcinoma(exceptthe2glandsfoundinthemiddleleftoftheimage)III.FEATUREEXTRACTION46imageswithlabeledregionsareusedfortheanalysis.Non-overlapping32x32pixelblocksaretakenfromthelabeledregions.Themethodsutilizedfortheextractionofprostatetexturefeaturesare:1)Gauss-Markovrandomfield(GMRF)model[8]-[9].2)Samplecorrelationfunction[8].3)Relativeentropymethod[10].GMRFmodelparametersareestimatedusingmaximumlikelihoodmethod.Neighborhoodsystemsbetween1and9areusedfortheextractionoffeatureset.Twodimensionaltimeseriesandrandomfieldmodelsprobablyarethemostcommonclassofimagemodels.Thedifferentclassofthesemodels,suchasautomodels,multi-levellogisticmodel,andhierarchicalGibbsmodel[11],characterizesthelocalstatisticalinformationintheimage.Incomputervision,imageisdefinedasa2Dspatiallydiscretearrayofintensityimagesamples,andfromthispointofview,theusefulnessofmodelingtheimagewitha2Drandomfieldisobvious[12].Samplecorrelationfunctiononthelattice,definedasthelos