人工智能核心550篇论文 医疗方向6.癌症 1707.08401v3.pdfVIP

Detectingandclassifyinglesionsinmammograms

withDeepLearning

˝1*´23´4´1

DezsoRibli,AnnaHorvath,ZsuzsaUnger,PeterPollner,andIstvanCsabai

1¨¨´

DepartmentofPhysicsofComplexSystems,EotvosLorandUniversity,Budapest

23rdDepartmentofInternalMedicine,SemmelweisUniversity,Budapest

3DepartmentofRadiology,SemmelweisUniversity,Budapest

4MTA-ELTEStatisticalandBiologicalPhysicsResearchGroup,HungarianAcademyofSciences,Budapest

*dkrib@caesar.elte.hu

7

1ABSTRACT

0

2

v

oInthelasttwodecadesComputerAidedDiagnostics(CAD)systemsweredevelopedtohelpradiologistsanalyzescreening

Nmammograms.ThebenefitsofcurrentCADtechnologiesappeartobecontradictoryandtheyshouldbeimprovedtobe

ultimatelyconsidereduseful.Since2012deepconvolutionalneuralnetworks(CNN)havebeenatremendoussuccessin

9imagerecognition,reachinghumanperformance.Thesemethodshavegreatlysurpassedthetraditionalapproaches,whichare

similartocurrentlyusedCADsolutions.DeepCNN-shavethepotentialtorevolutionizemedicalimageanalysis.Wepropose

]

VaCADsystembasedononeofthemostsuccessfulobjectdetectionframeworks,FasterR-CNN.Thesystemdetectsand

classifiesmalignantorbenignlesionsonamammogramwithoutanyhumanintervention.Theproposedmethodsetsthestate

CoftheartclassificationperformanceonthepublicINbreastdatabase,AUC095.Theapproachdescribedherehasachieved

s.the2ndplaceintheDigitalMammographyDREAMChallengewithAUC085.Whenusedasadetector,thesystemreaches

c