外文资料-- Diagnosis Technology Research Of Mammographic Masses in Content-based Image Retrieval.PDF
DiagnosisTechnologyResearchOfMammographicMassesinContent-basedImageRetrievalSongLi-xin,WangQing-yanCollegeofElectrical&ElectronicEngineeringHarbinUniversityofScienceandTechnologyHarbin,Chinalixins99yahoo.com.cnWangLiDept.ofgynecologyHeilongjianghospitalHarbin,Chinawangli1439163.comwangqingyan1984163.comAbstractInordertoassistdoctortodiagnosisofmammo-graphicmasses,amethodisproposed.22featuresareextractedfromeachqueriedregionofinterest(ROI).Ak-nearestneighbor(KNN)algorithmisusedtoretrievesimilarimagesfromdatabase,andfurthercalculatethemutualinformation(MI)betweenthequeriedimageandtheimageswhichareintheretrievalresults,soastoimprovetheretrievalperformance.Finally,theschemetakesthefirstnineimageswiththehighestMIscoresasthefinalretrievalresults.Withthepurposeofprovidingavailabledecision-makinginformationofdiagnosticaids,wecompareandanalyzethreecalculatingmethodsofdecisionindex.TheExperimentresultsshowthatthemethodisbetterthanmethodofusingKNNonly,andimprovetheaccuracyofdiagnosiseffectively.Keywords-mammographyimage;content-basedimagere-trieval;computer-aideddiagnosis;mutualinformationI.INTRODUCTIONScreeningmammographyisconsideredthemostreliableandeffectivemethodintheearlydetectionofbreastcancer.However,mammographyimageisblurryanditscontrastislowduetoimagingprinciple,manysmallerlesionsarenoteasytobeobservedandextractedowingtotheoneswhichhavebeenoverwhelmedbythenormalbreasttissue.Therefore,earlydetectionanddiagnosisofbreastcancerreliedheavilyontheradiologistssubjectiveviews.Computer-AidedDetectiontechnologyprovidesanvaluable“secondreview”1forradiologistsintheearlydetectionprocessofbreastcancer.ThetraditionalCADsystemofbreastmassesdetectiongenerallybasedonartificialneuralnetwork.ThemethodofContent-BasedImageRetrievalnotonlyeliminatedtheburdenofdesigningandtrainingneuralnetworkclassifiers,butalsomadefulluseofpastdiagnosisofbreastmassesoftheoriginalempiricaldata.Accordinglyitcaneffectivelyhelpphysiciansimprovetheaccuracyoftumordiagnosis.AlthoughCBIRhasbeenappliedinanumberoffields,thebreastimageswhicharecharacterizedbyhigherresolution,biggerimagesimilarityandmoreinformationaredifferentfromothers.Itisdifficulttoretrievesimilarimagewiththesamepathologicalcharacteristicsfromimagelibrary.Itisthereforeessentialtoestablisharetrievalsystemwithmedicalpractice.MammographyofCBIRisstillastudyingdirection.Someresearchinstituteshavebeencarriedoutrelevantresearch.BinZheng2hasproposedaninteractiveretrievalmethodwhichimprovedthevisualsimilaritybetweentheresultimagesandthequeriedimages.GeorgiaD.Tourassi3adoptedmutualinformationasthesimilaritymeasurementamongimages,classifiedtheregionofinterestbythemostsimilarimagesofretrieval.HilaryAlto4studiedthevariouscombinationsofshape,textureandedgesharpnessfeatures,whichusedinclassifyingbenignandmalignanttumormasses.Toimprovetheretrievalaccuracyandachievethediagnosis,thispaperstudiesacombinationofKNN+MIretrievalmethodofbreastmassesdetection,andgivesadecisionindexofaideddiagnosisonthebasisofthisstudy.II.CONTENT-BASEDMAMMOGRAPHICMASSESIMAGERETRIEVALAftersubmittingaqueriedROIimage,thesystemwillautomaticallyextract22featherswhichmatchthefeathersinthedatabase,obtainseveralfrontimagesaccordingtothesimilarityfromhightolow,andfinallycalculatethedecisionindexandanalyzethepathologicalinformationofmassesaccordingtotheretrievalimagesandtheirpathologicalinformation.TheoverallprocessofthemethodisshowninFig1.Figure1.ThediagramoftheoverallprocessFig.2showstheresultoftheretrieval.ThequeriedROIisintheupper-leftandthepathologicalinformationisthefollowing.ThefirstnineROIsareintheleftandthelettersaboveeachROImeanclasses,“M”:malignant,“B”:benign.A.FeatureExtractionTheconventionalcontent-basedretrievalaimedatimprovingvisualsimilaritybetweentheretrievalimagesandthequeriedimage,butthebreastX-rayimagesarevisuallyverysimilar,sofeatureselectionprocessshouldnotbebasedsolelyonthevisualsenseofthesimilarities.Doctorsalsotend978-1-4244-4713-8/10/$25.00©2010IEEEtolikethesamekindofimagesasthesimilarimages,whichisthesimilarityinthemedicalsense.Therefore,thisfeatureselectionbasedonthefollowingprinciples:ifafeatureisvalidinclassification,itisalsovalidinretrieval.Thus,themorewhicharethesamewiththequeriedimagesintheresults,themoreeffectivetheretrievalis.Figure2.TheresultoftheretrievalAfterthequeryROItobeobtained,iftocalculatetherelevantcharacteristicsofasuspiciouslump,thenitrequiressegmentationofsuspiciousmasses.Thesegmentationofsuspiciousmassesisdividedintothreestepsinthistest:1.Removethe“backgroundtrend”5ofthequeriedROIimage;2.Restraintheadjacenttissuesofsuspiciousmasses;3.Segmentthesuspiciousmasswiththeimproved"MultilayerTopographicSegmentation"6.Afterthat,22featuresareextractedfromeachROIasthefeatherset,includingBinZhengsetal2twelvefeathers,NicholasPetricks7sevenfeathersandRenchaoJins8threefeathers.B.SimilarityMeasurementTheimagesimilaritymeasureisthesimilaritybetweentheimagefeatures.Thesimilaritymeasurementmethodswillhaveadirectimpactontheperformanceofimageretrieval.ThispaperproposesacombinationofKNNwithMIsimilaritymatchingalgorithm.1)K-nearestneighboralgorithm:Amulti-featurek-nearestneighbor(KNN)basedalgorithmwasappliedtosearchforthe“computationallysimilar”ROIsinthereferencelibrary.Similaritywasmeasuredbythedifferenceinfeaturevalues,()rifxbetweenaqueried()qROIyandareference()iROIsxinamultidimensional(n)featurespace,21(,)()()nqirqrirdyxfyfx=(1)Thesmallerthedifference(“distance”),thehigherthedegreeofthecomputed“similarity”isbetweenanytwocomparedregions.Thecomputeddistancesbetweenatest(queried)regionandeachofthestoredreferenceregionsweresorted(rankordered)fromthesmallesttothelargest.ThefirstKregionsinthelistwerethenselectedastheK“mostsimilar”(orthebest“matched”)referenceregions.Adistanceweightwasdefinedas020201,(,)1,qiidddyxwddd>=Andtheclassificationscore,ortheprobabilitythataregionisactuallymalignant,wascomputedas111MiiMNijijwPww=+,KMN=+(2)WhereNisthenumberofmalignantmassregionsandMisthenumberofbenignmassregionsthatwereselectedinthesetofK“mostsimilar”ROIs.2)Maximummutualinformationmethod:TheresultsretrievedbyKNNmethodarefurthermatchedbythemaximummutualinformationmethod,andthefinalresultisbetterthanothersbyusingKNNonly.Thecorrelationbetweentworandomvariablesentropy,isalsoknownasmutualinformation.Mutualinformationbetweentworandomvariablescanserveasastatisticalmeasureofcorrelation.Inthepreviousstudy,imageretrieval,mutualinformationalsohasbeenappliedtocontent-basedmedicalimageretrieval,andhasachievedrelativelygoodresults.GiventwoimagesXandY,theirMII(X;Y)isexpressedas:2(,)(,)(,)log()()xyxyxyxyPxyIxyPxyPxPy=(3)WherePXY(x,y)isthejointprobabilitydensityfunction(PDF)ofthetwoimagesbasedontheircorrespondingpixelvalues.PX(x)andPY(y)arethemarginalPDFs.Thebasicideaisthatwhentwoimagesarealike,themoreinformationXprovidesforYandviceversa.Therefore,theMIcanbethoughtasanintensity-basedmeasureofimagessimilarity.IfthequeryimageXandastoredimageYdepictsimilarstructures,thenthepixelvalueinimageXshouldbeagoodpredictorofthepixelvalueatthecorrespondinglocationinimageY.Consequently,theirMIshouldbehigh.AsshowninEq.3,theMIestimationoftwomammo-gramphicROIsrequirescomputationofthejointandmarginalPDFs.Wefollowedthehistogramapproach9forthetask.SincetheimagesofDigitalDatabaseforScreeningMammography(DDSM)consideredinourstudyare12-bitimages,thePDFswereestimatedusingareducednumberof256equal-sizedintensitybinstoavoidpotentialoverestimationerrors10.ThisisatypicalpracticeforMIestimationinimageregistration.C.DecisionIndexBesidestheretrievedROIimages,thedecisionindex(DI)indicatingtherelativeprobabilitythataROIcontainsamasscanbecalculatedautomaticallywithaformulaandoutputtotheuser.AhigherDIvaluemeansahigherprobabilitythattheROIcontainsamass.TheformulaforcalculatingtheDIisbasedonthemethodsproposedbyGeorgiaD.Tourassietal3.1211(,)(1()()(,)(1()(,)(1()MQIIIQMNQIIQJJIJSYXKRXDIYSYXKRXSYXKRX=×+=×+×+(4)WhereMisthenumberofimagesretrievedfromthedatabasethosecontainmassROIs.NisthenumberofimagesretrievedfromthedatabasethosecontainnormalROIs.K=M+N.Rank(XI)istheorderingnumberofXIwhentheretrievedROIimagesaresortedindescendingorder.Itcanbeseenthatforeithermethod,thehigherDImeansahigherprobabilitythattheROIcontainsamass.DI2consi-deredthefactoroforderingnumberofXIandassignedarighttoeachofsimilaritymeasurevaluesanditgivesabetterper-formanceinourevaluationexperiments,sowetakeitasourinitialdecisionindex.III.EXPERIMENTALRESULTSANDANALYSISROIintheimagedatabasecomesfromDDSMofUniversityofSouthFlorida.TheROIdatabaseincludes514malignantROIsand321benignROIs.EachROIisillustratedinanimagewithsizeof125×125pixels.Thedepthofimageis12bits.EachROIcontainsatmostonemass.NomassintheROIweextractedisonthechestwall.Recallrateandprecisionrateisthestandardinformationretrievalevaluationmethod.ThenumberofimagesreturnedKtakesaveryimportanteffectfortheperformanceofKNNretrievalsystem.TheaverageprecisionratewillbecalculatedtoobtainanoptimalKvalue.FromFig.3itisconcludedthatprecisionrateisnotmuchdifferentforthedifferentKvalues.However,consideringthemutualinformationmatchingfortheresults,therelativelysmallandtheprecisionvalueslightlyhighKisselected,K=25.Figure3.TheaverageprecisionofdifferentKvaluesAthresholdvalueisusedasadividingpointbetweenbenignandmalignantmasses,andtheFig.4showsthedistributionofbenignandmalignantmassesofdecisionvaluesinthedatabase.Thereisnocleardemarcationpointbetweenthemassesfromthehistogram.Thus,athresholdvalueshouldbedefinedbetween0and1.The“ERRORRATE”isdefinedbythenumberofmalignantROIofDITandthenumberofbenignROIofDIT,accordinglybythenumberofwrongdecisionunderT.ItischangingwithT,andcanbefoundoutbyExhaustiveAttackmethod.Figure4.ThehistogramofdecisionindexThenareceiveroperatingcharacteristic(ROC)curvecanbeplottedtoevaluatetheperformanceofusingoursystemtoclassifybetweentrue-positiveandfalse-positivemassregions.TheareaunderROCcurve(AUCvalue)isusedastheindexofperformance.Theleave-one-out11samplingschemeandtheROCcurveanalysisareusedfortheassessmentofoursystem.EachtimeaROIimageischosenfromthedatabaseFigure5.TheROCcurveoftwomethodsindifferentDiesasthequeryROIimage,thentherestROIimagesformatestdata-base.Theprocedureisperformedrepeatedly,eachtimeaROIimageinthedatabaseischosenasaqueryimage.Fig.5showedtheROCcurveoftwomethodsindifferentDies.Tab.IshowedtheAZvaluesoftwomethodsindifferentDies.Analysingandcomparingthisthreemethods:DI1addssomeeffectsofsimilarityonthebasisofsequence.PandDI2arebasedonweightedsumofthereciprocalsquareofthedistanceasthemainbasisforsimilaritymatching.DI2addedtheweightedsumofsequencesimilarityplaysacertainroleinimprovingdetectionaccuracyrate.ItcanbeshowedfromTab.IthatDI2isobviouslybetterthantheothers,themethodofKNN+MIisbetterthanmethodofusingKNNonly.TABLEI.THEAZVALUESOFTWOMETHODSINDIFFERENTDIESDI1PDI2KNN0.7084±0.01830.7353±0.01740.7452±0.0171KNN+MI0.7198±0.01780.7612±0.01660.7984±0.0154EachmammographicimageinDDSMdatabasecontainspathologydiagnosticinformationoflocationofalesion,whichisgivenbyseveralradiologistsbasedontheirexperiencesofyearsinthediagnosis.Then,intheROIresults,weextractedpathologyinformationforeachROIrespectivelyfromthepathologymessagedata,includingDensity,Shape,Margins,Assessment,Subtlety,etc.Tab.IIshowsthatthecomparativeprecisionineachpathologicalinformationintwomethods.TABLEII.THEPRECISIONOFEVERYPATHOLOGYINFORMATIONINTWOMETHODSDensityShapeMarginsAssessmentSubtletyKNN62.15%70.58%71.46%67.25%72.53%KNN+MI65.48%76.97%78.65%61.36%69.57%ItisclearfromthesedatathatthereisacertaindifferenceButingeneralnotintwodifferentretrievalmethodsforthedifferentpathologicalinformation.Therefore,themethodofKNN+MIcanbringagreatreferencevalueforadoctorinthepathologicaldiagnosisandprovideagreathelptotheearlydiagnosisofbreastmasslesions.IV.CONCLUSIONSAmethodforcomputer-aideddetection(CAD)ofmammographicmassesisproposedandaprototypeCADsystemispresented.ThesystemcanautomaticallyevaluatethepossibilitythataROIismalignantorbenignbyretrievingsimilarROIimagesfromthedatabaseandcalculatingtheDIvalueforeachROI.Thesystemperformanceisevaluatedusingtheleave-one-outsamplingschemeandROCcurveanalysismethodbasedontheDIsoutputbytheprototypesystem.CBIR-basedCADisausefulmethodforcomputer-aideddetectionofmammograhpicmasses.ACKNOWLEDGMENTSTheworkwassupportedbyNaturalScienceFoundationofHeilongJiangProvince(F200912).REFERENCESS1L.J.WarrenBurhenne,S.A.Wood,C.J.DOrsi,S.A.Feig,D.B.Kopans,etal,“Potentialcontributionofcomputer-aideddetectiontothesensitivityofscreeningmammography,”Radiology,vol.215,no.2,pp.554-562,2000.2Z.Bin,A.Lu,L.A.Hardesty,J.H.Sumkin,C.M.Hakim,etal.“Amethodtoimprovevisualsimilarityofbreastmassesforaninteractivecomputer-aideddiagnosisenvironment,”.MedicalPhysics,vol.33,no.1,pp.111117,2006.3G.D.Tourassi,R.Vargas-Voracek,C.E.Floyd.“Content-basedimageretrievalasacomputeraidforthedetectionofmammographicmasses,”In:ProceedingsofSPIE,5032,pp.590597,2003.4H.Alto,R.M.Rangayyan,J.E.L.Desautels.“Content-basedretrievalandanalysisofmammographicmasses,”JournalofElectronicImaging,vol.14,no.2,pp.023016117,2005.5M.Aoyama,Q.Li,S.Katsuragawa,F.Li,S.Sone,etal.”Computerizedschemefordeterminationofthelikelihoodmeasureofmalignancyforpulmonarynodulesonlow-doseCTimages,”MedicalPhysics,vol.30,no,3,pp.387394,2003.6B.Zheng,Y.H.Chang,D.Gur.”ComputerizedDetectionofMassesinDigitizedMammogramsUsingSingle-ImageSegmentationandaMultilayerTopographicFeatureAnalysis,”AcadRadiol,vol.2,no.11,pp.959966,1995.7N.Petrick,H.P.Chan,D.Wei,B.Sahinet,M.A.Helvie,etal.“Automateddetectionofbreastmassesonmammogramsusingadaptivecontrastenhancementandtextureclassification,”MedicalPhysics,vol.23,no.10,pp.16851695,1996.8R.C.Jin,B.Meng,E.M.Song,X.Y.Xu,L.Jiang,“Computer-aidedDetectionofMammographicMassesBasedonContent-basedImageRetrieval”,Proc.ofSPIEVol.6514,65141W,2007.9F.Maes,A.Collignon,D.Vandermeulen,G.Marchal,G.Suetens,"Multimodalityimageregistrationbymaximizationofmutualinformation,"IEEETrans.Med.Imag.,vol.16,pp.187198,1997.10A.Treves,A.Panzeri,“Theupwardbiasinmeasuresofinformationderivedfromlimiteddatasamples,”NeuralComputation,vol.7,pp.399407,1995.11B.Efron,R.J.Tibshirani,AnIntroductiontotheBootstrapM.NewYork:CRCpress,pp.141150,1993.