内镜下图像分类器在胃肠道恶性肿瘤诊断中的应用(IJIGSP-V9-N6-6)

I.J.Image,GraphicsandSignalProcessing,2017,6,45-54PublishedOnlineJune2017inMECS()DOI:10.5815/ijigsp.2017.06.06Copyright©2017MECSI.J.Image,GraphicsandSignalProcessing,2017,6,45-54ImageClassifiersinEndoscopyforDetectionofMalignancyinGastroIntestinalTractKVMahendraPrashanth1,VaniV21Professor,Dept.ofElectronics&CommunicationEngineering2ResearchScholar,Dept.ofElectronics&CommunicationEngineeringSJBInstituteofTechnology,VisvesvarayaTechnologicalUniversity,Bangalore,IndiaEmail:[e-mail],[e-mail]—WirelessCapsuleEndoscopy(WCE)isoneofthemethodsforexaminationofgastrointestinal(GI)disorderssuchasobscureGIbleeding,Crohnsdisease,polypsetc.WCEhasbeenrecognizedasalessexpensiveandpainlessprocedureforthediagnosisofGItract.ThispaperexaminesthevariousimageclassifiersdesignedanddevelopedforthepurposeofendoscopyfocusingspecificallyonWCE.ItisrevealedthatdesigningasuitableimageclassifierisanimportantprerequisiteforaccurateandprecisediagnosisofmalignancyinWCE.TheassessmentonvariousimageclassifiersusedforthediagnosisofpathologiesindifferentpartsofGItractshowsthatclassifiershavereducedthediagnosistimeformedicalexpertsandalsoprovidedreasonablyaccuratediagnosisofmalignancy.However,correlatingclassifiersandrelatedpathologiesisstillobservedtobechallenging.Inviewofthefactthatearlydetectionmaydecreasethemortalityratesignificantly,inclinationtowardscomputeraideddiagnosisareexpectedtoincreaseinfuture.Thereisaneedforadvancedresearchinthedevelopmentofarobustcomputeraideddiagnosissystem,capableofdiagnosisofvariouspathologiesinGItractwithhigherdegreeofaccuracyandreliability.Further,thestudydepictsthatadirectcomparisonofresultsofclassifiersuchasaccuracy,prediction,sensitivity,specificityandprecisiontoevaluateitsperformanceischallengingduetodiversityofimagedatabases.Moreresearchisneededtoidentifyandreducetheuncertaintiesintheapplicationofimageclassifiertoimprovethediagnosisaccuracy.IndexTerms—Imageclassification,WirelessCapsuleEndoscopy(WCE),MachineLearning,SupportVectorMachine(SVM)I.INTRODUCTIONWirelessCapsuleEndoscopyisoneofthestandardimagingtechniquestodayforexaminationofgastrointestinaltract(GItract),especiallythesmallintestine;whichisnoteasilyreachablebynormalupperendoscopyandcolonoscopymethods[1].Byconventionalendoscopictechniques,onlytheproximalduodenumanddistalileumofthesmallintestinecouldbeexamined.Thediscoveryofcapsuleendoscopeledtothenewtechnologybymeansofwhichevenasmallbowelbecamevisibletothegastroenterologist[2].TheswallowablewirelesscapsuleendoscopytakesimagesoftheGItractwhicharelaterdownloadedforanalysisinacomputer[3].Conventionalmethodofdiagnosisinvolvesvisualanalysisofimagesbyanexpertphysician.Computer-aideddiagnosisincludesextractionoffeaturesoftheimagesandautomaticclassificationofimagesindicatingthepresenceorabsenceofmalignancy;thereby,easingthetaskofobjectiveinterpretationbyaphysicianinendoscopy.Awirelesscapsuleprovidedcontinuousvideostreamoftheinnermucosalumentubularstructure.Typically,inWCEtheimageexaminationforeachpatientconsistsof8hoursofvideowhichisaround55,000frames.Theprocedureinvolvingexaminationandreviewingthevideofromcapsuleendoscopyhasbecomecumbersome;asitrequiresconcentrationforlongdurations.Computerizedimageanalysisalgorithmscanreducethetimerequiredtoreviewforimageexaminationbyexpertsandcanaugmentthedecisionmakingprocesses[4].Thus,automaticCADmethodsarerequiredwhichcanhelpinanalysisanddiagnosis.SignificantresearchhasbeenobservedfromthelastdecadeandthenumberofpublicationsinareaofdesignanddevelopmentofcomputeraideddiagnosticsystemsforWCEhassubstantiallyincreasedto35%startingfromyear2004[5].TheobjectiveofthispaperistoprovideacomprehensiveoverviewofthelatestadvancementsandtechniquesusedforimprovingclassificationaccuracyinWCE.Further,theissuesandlimitationsinexistingimageclassificationtechniques;whicharebeingusedinWCEarediscussed.Possibilitiestoimprovetheimageclassifiers,providingafaster,moreaccurateandprecisediagnosisofmalignancyinGItractarebeingdeliberated.Literaturehasbeenreviewedonkeyingthekeywords:capsuleendoscopy,histogram,classifierlearning,computeraideddiagnosis,medicalimagedata,imageclassifierinvariousprevalentpublicationssuchasSpringer,IEEE,SAGE,Elsevier,PubMed,ACM,Hindawi,ScienceDirectwhichhavebeenpublishedduring2000-2015.Section2discussesthepathologiesunderinvestigationinGItract.DevelopmentofbetterimageclassifiersisdescribedinSection3.Section4providesacomprehensivestudyofthegeneralclassificationofImageClassifiersandImageClassifiersemployedforcapsuleendoscopy.Section5describestheevaluationof46ImageClassifiersinEndoscopyforDetectionofMalignancyinGastroIntestinalTractCopyright©2017MECSI.J.Image,GraphicsandSignalProcessing,2017,6,45-54classifierperformance.AdvantagesandchallengesinclassifiersarediscussedinSection6.Section7depictstheoverallconclusionofthestudy.II.PATHOLOGIESUNDERINVESTIGATIONINGITRACTUSINGWCEItisobservedthatavarietyofpathologies;thataretargetedbydifferentComputerAidedDiagnosticsystemsfordetectionandclassificationinGItractdoexist.Someofth