(a)(b)Figure1.36Thicksectionimagesof(a)invasivecarcinoma,and(b)radials - PDF document

32C (a)(b)Figure1.36Thicksectionimagesof(a)invasivecarcinoma,and(b)radialscar. (b)(c)Figure1.37Spiculatedtumormassisthemostusualmanifestationofinvasivecarcinomas.(a)Macroscopicimage,(b)radiologicimage,and(c)histologicimage. UBGROSSMORPHOLOGYOFALTEREDBREASTTISSUEInvasionmostoftenevokesproliferationof“broblastsandmyo“broblastsaswellascollagenizationinthestromadamagedbythetumorcells„areactioncalledstromaldesmoplasia,whichpreventsthetumorcellsfrominvadingthestromaevenlyinalldirections.Thisprocessleadstothedevelopmentofaspiculated(stellateŽ)tumormass,themostusualmanifestationofinvasivecarcinoma.Fig-ure1.36(a),demonstratesa3Dimageofsuchatumorandcomparesittothe3Dimageofaradialscar[Fig.1.36(b)].Thetypicalmacroscopicappearanceofastel-latecarcinomaisillustratedinFig1.37.Abouttwo-thirdsofinvasivecarcinomasareofstellateshape.Theshapeoftheinvasivetumorisrelatedtoitshistologictype.Thestellatecarcinomasaremostofteninvasivelobular,tubular,orgradeIorgradeIIinvasiveductalnototherwisespeci“ed(NOS)cancers(seeDiagram1.5);thesetogethercorrespondtomorethan80%ofstellatetumors.Benignstellatelesionsarerare.Onlytheappearanceofelastic“bersinthestroma(asinradialscars)orcollagenizationduetoscarformationleadstolesionsimitatingthestellateshapeofbreastcarcinoma.Thus,astellatelesiononamam-mogramismostoften(inmorethan90%ofcases)malignantandwarrantsfur-therdiagnosticandtherapeuticinterventions.Althoughspiculated,radialscarslackawell-formedtumorbody[Fig.1.38(a)]andappearasblackstarsŽonamammogram[Fig.1.38(b)]„incontrasttothewhitestarsŽincarcinomacases[Fig.1.37(b)].Thehistologicappearanceisalsotypicalastheyexhibitanelasticcoreandacoronaofhyperplasticchanges[Fig.1.38(c)].Inaminorityofcases,radialscarsmaycontainlow-gradeinsitucarcinomafociintheircoronaandrep-resentaspecialtypeŽofinsitucancer.Ifthedesmoplasticstromalreactiononinvasionofthetumorcellsisweakorabsent,thecellsmayinvadeevenlyinalldirections,givingrisetoacircular(oroval)tumormass.ThesetumormassesaresimilartobenignormalignantlesionsDiagram1.5Distributionof910consecutivestellateinvasivecarcinomasbytumortype,Falun1996…2003. 9.1DETECTIONOFICROCALCIFICATIONSACKGROUNDANDOTIVATIONInWesterncountries,womenhaveahigherthan1-in-8chanceofdevelopingbreastcancerduringtheirlives.Breastcancerrepresentsthemostfrequentlydiagnosedcancerinwomen.TheNationalCancerInstituteofU.S.A.estimatesthat,basedoncurrentrates,13.2%ofwomenborntodaywillbediagnosedwithbreastcanceratsometimeintheirlives.Inordertoreducemortality,earlydetectionofbreastcancerisimportant,be-causetherapeuticactionsaremorelikelytobesuccessfulintheearlystagesofthedisease.Forwomenwhosetumorswerediscoveredearlybymammography,the“ve-yearsurvivalratewasabout82%asopposedto60%forthecaseswherethetumorswerenotfoundearly.Mammographyiscurrentlythebestradiologicaltechniqueavailableforearlydetectionofnonpalpablebreastcancer.However,itisdif“cultforradiologiststoprovidebothaccurateanduniformevaluationsforthelargenumberofmammo-gramsthattheyhavetointerpretinscreeningprogramswheremostofthecasesarenormal;ithasbeenobservedthat10…30%ofbreastlesionsaremissedduringroutinescreening.Thesituationisevenmorechallengingsincetheearlymalig-nancieshavesmallsizeandsubtlecontrastwhencomparedwithnormalbreaststructures.Doublereading(ascarriedout,forexample,bytworadiologists)helpstoreducethenumberoffalsenegativesby5…15%.Digitalimage-processingtechniquesrepresentusefultoolsforhelpingradiol-ogiststoimprovetheirdiagnosiswiththeaidofcomputersystems.Inthissense,differentCAD(computer-aideddiagnosis)toolshavebeendevelopedforimprov-ingimagequality,identifyingmalignantsigns,enhancingmammographicfeatures,etc.Ontheaverage,thereaderssensitivitycanbeincreasedby10%withtheassis-tanceofCADsystems.SomeworkshavestudiedthispotentialofCADtoimproveradiologistsperformanceindetectingclusteredmicrocalci“cations.Thereareanumberofdifferentclassesofabnormalitythatmaybeobservedinmammograms.Oneofthemostsigni“canttypesofmammographicabnormal-ityismicrocalci“cation.Microcalci“cationsaretinygranulelikedepositsofcal-cium.Theyarerelativelybright(dense)incomparisonwiththesurroundingnormalandareuptoabout1mmindiameter,withanaveragediameterof0.3mm.Microcalci“cationsareofparticularclinicalsigni“cancewhenfoundinclustersofthreeormorewithinasquare-centimeterregionofamammogram.Lanyidescribedmicrocalci“cationsasthemostimportantleadingsymptominmam-mographicdetectionofpreclinicalcarcinomas.ŽSicklesnotedthatmorethan50%ofnonpalpablecancershadmammographicallyvisiblecalci“cations,andin36%ofnonpalpablecancers,calci“cationsweretheonlysignofabnormality.Inanimportantstudyofcancersmissedinscreeningmammography,itwasobservedthatthepresenceofmicrocalci“cationswasthepredominantfeaturein18%ofthemissedcancers. approachconsistsofusingtheanatomicknowledgeofthestromatodevelopaconstraintbetweenthecon“rmedportionoftheskinlineandthestromaedge.Thisstepusesthedensityconcepttodevelopthethreshold.Thestromaedgeissmoothedby“ttingaspline,andconstraintsareestablishedbetweentheinitialcon“rmedportionoftheskinlineandthesmoothedstromaedge.Theconstraintsarethenpropagatedtowardtheupperandlowerbreastzonestocorrecttheweakareasoftheboundary,therebygettingclosertothetrueedge.Thepropagationisperformedbyusingagreedyapproachincombinationwithanatomicalconstraints.Comparingtheresultsofthealgorithmwiththoseofoneofthebestmethodsreported,namelythedeformablemodelofFerrarietal.,animprovementinskin-lineestimationisdemonstratedbythisapproach.Notethattheground-truth(GT)boundariesusedforthecomparisonwerepreparedandtracedbyradiologists.Thechapterisorganizedasfollows.Section11.2presentstheoveralldesignofthesystem.Section11.3providesadescriptionofadaptivethresholdingtoobtainaninitialskinline.InSection11.4,ashapeanalysismethodfortheextractionoftheinitialcon“rmedportionoftheskinlineispresented.InSection11.5,themethodsfortheextractionoftheedgeofthestromaandspline“ttingarepresented.Sec-tion11.6presentstheproposeddependencyapproachtoobtainthe“nalbreastskinline.SeveraltechniquesandmetricsforperformanceevaluationandcomparisonhavebeenimplementedasdescribedinSection11.7,includingthepolylinedis-tancemeasure(PDM)andtheHausdorffdistancemeasure(HDM).TheresultsoftheanalysisareshowninSection11.8,andthepaperisconcludedinSection11.9.11.2ABVERVIEWOFTHEROPOSEDTheconceptofthealgorithmarisesfromthefactthattheedgeofthestromaofthebreasthasausuallyuniformdistancetothebreastskinline(seeFig.11.3).Thestromaedgeisnotaffectedbythesystemnoisebecauseofthehighcontrastbetweenthebright(dense)stromaregionandthelow-intensity(low-density)fattyperipheralregion.Thebackgroundregionandthebreastregionaredistinguishedbyadiscontinuityinthehistogram.Thisdiscontinuitymaybeestimatedusinganadaptivethresholdingtechnique,suchasthemethodproposedbyOjalaetal.stepisrepresentedinFig.11.2asblock(A).Becausethestromaisahigh-densityzone,itcaneasilybemaskedoutbyabi-modalitythresholdingprocedure.Onesuchthresholdingprocedurehasbeenpro-posedbyOtsu,whichiswidelyacceptedinmedicalimageprocessing.TheaddedadvantageofOtsusprocedureforthresholdcomputationisitssimplicity.Thisstepisidenti“edinFig.11.2asblock(B).Therefore,aninitialestimateoftheouter-andinner-edgeofthebreastisobtainedusingtwodifferentthresholdingprocesses,yieldinginitialestimatesoftheskin-lineboundaryandthestromaedge.ThisisshownbytheparallelpathsinFig.11.2.Havingextractedaninitialcon“rmedportionoftheskinlinefromtheinitialskin-lineestimateandtheaccuratestromaedgeinformation,anovelstrategytopropagateorextrapolatethepartialboundarytoafullskin-lineboundarybased IAGNOSISINMAGINGTheexcessivebiopsyofbenignlesionsraisesthecostofmammo-graphicscreeningandresultsinemotionalandphysicalburdentohundredsofthousandsofpatientseveryyear,aswellas“nancialburdentosociety.Thus,itwouldbeexceedinglyvaluabletoproducecomputer-aideddiagnosis(CADx)sys-temsthatcouldaidinthedecisiontorecommendbiopsyorrecommendshort-termfollow-upmammography.Thoseverylikelybenigncasesmaybemanagedwithshort-termfollow-up,whichwillavoidahugenumberofunnecessarybiop-sieswhilemaintainingtheveryhighsensitivityofcancerdetection.Moreover,sinceitisestimatedthatabouthalfofmissedcancersaremissedduetomisinter-pretationratherthanoversight,itmaybepossibletoincreasethesensitivityofmammographythroughCADx.27.1.3BRIEFHISTORYOFCADxinbreastimagingstartedinearnestintheearly1990s,andtherehavelit-erallybeenhundredsofpublicationsintheliterature.MostofthemajorpapershavebeeninMedicalPhysicsAcademicRadiology,withoccasionalclinicalevaluationsinRadiologyManyreviewsofthe“eldhavebeenwritten.Amongrecenteffortsinpar-ticular,GigerprovidesabroadoverviewofthestateoftheartinCADandCADx,andSampatetal.provideadetailedsummaryofthemanydifferentapproachesem-ployedandperformanceresultsattained.Inthefollowingsections,citationswillbeprovidedtootherrepresentativeand/orrecentpublicationsinthecontextofthemajorchallengesstillfacedinCADxresearch.27.2CADLASSIFIERThereisnouniversalstructureforCADxsystems,butmostdofollowthegen-eral”owchartthatisverysimilartothatofCADsystems(seeFig.27.1).Inthisexample,thesystemallowsforinputsfromthreedifferentsources:computer-extractedimage-processingfeatures,radiologist-interpreted“ndings,andpatienthistory“ndings.Thevariousfeaturesareselectedaccordingtosomerationale,thenmergedtogetherusingaclassi“er.Eachofthesestepsareexplainedinfurtherdetailinthefollowingsections.27.2.1LINEARVERSUSNONLINEARCLASSIFIERSMuchoftheearlyworkinCADxwasbasedonthearti“cialneuralnetwork(ANN)nonlinearclassi“er,whichisdescribedinmanyclassictexts.Neuralnetworkswerequicklyembracedbythe“eld,andthishistoryisreviewedbelowtomakeanimportantcautionarynote.AnexampleofthemodelarchitectureisshowninFig.27.2.