资源描述：
《A reinforcement agent for object segmentation in ultrasound images .pdf》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。
1、Availableonlineatwww.sciencedirect.comExpertSystemswithApplicationsExpertSystemswithApplications35(2008)772–780www.elsevier.com/locate/eswaAreinforcementagentforobjectsegmentationinultrasoundimagesa,c,*a,cbFarhangSahba,HamidR.Tizhoosh,MagdyM.M.A.SalamaaDepartmentofSystemsDesignEng
2、ineering,UniversityofWaterloo,Waterloo,CanadabDepartmentofElectricalandComputerEngineering,UniversityofWaterloo,Waterloo,CanadacPatternAnalysisandMachineIntelligenceLaboratory(PAMI),UniversityofWaterloo,Waterloo,CanadaAbstractTheprincipalcontributionofthisworkistodesignageneralfra
3、meworkforanintelligentsystemtoextractoneobjectofinterestfromultrasoundimages.Thissystemisbasedonreinforcementlearning.Theinputimageisdividedintoseveralsub-images,andtheproposedsystemfindstheappropriatelocalvaluesforeachofthemsothatitcanextracttheobjectofinterest.Theagentusessomeima
4、gesandtheirground-truth(manuallysegmented)versiontolearnfrom.Arewardfunctionisemployedtomeasurethesimilaritiesbetweentheoutputandthemanuallysegmentedimages,andtoprovidefeedbacktotheagent.TheinformationobtainedcanbeusedasvaluableknowledgestoredintheQ-matrix.Theagentcanthenusethiskn
5、owledgefornewinputimages.Theexperimentalresultsforprostatesegmentationintrans-rectalultrasoundimagesshowhighpotentialofthisapproachinthefieldofultrasoundimagesegmentation.Ó2007ElsevierLtd.Allrightsreserved.Keywords:Reinforcementlearning;Imagesegmentation;Ultrasoundimaging1.Introduc
6、tionAmongdifferentmedicalimagemodalities,ultrasoundimaginghasaverywidespreadclinicaluse(CancerFactsMedicalimagesegmentationplaysaninvaluablerolein&Figures,2002).Inmanyapplications,weneedtoseg-manyapplicationsthroughthedemonstrationofanatomi-mentanobjectintheseimages.Butthecasesofte
7、nfacedif-calstructures.Segmentationisachievedbasedonsomeficultiesduetosomefactorssuchaspoorimagecontrastmeasurablefeaturessuchasintensity,textureandgradient.betweenareasofinterest,noiseandmissingordiffuseTechniquesusedinsegmentationsalgorithmsarehighlyboundaries.Furthercomplications
8、arisewhenthequalitydependentonthe
