一种病理图像自动标注的机器学习方法.pdfVIP

计算机研 究 与发展 DOI：10．7544／issn1000—1239．2015 JournalofComputerResearchandDevelopment 52(9)：2135—2144，2015 一 种病理图像 自动标注的机器学习方法 张 钢 钟 灵 黄永慧 (广东工业大学 自动化学院 广州 510006) (ipx@gdut．edu．cn) A M achineLearningM ethod forHistOpath0J0gicaJIm ageAutomaticAnnotation ZhangGang，ZhongLing，andHuangYonghui (SchoolofAutomation，GuangdongUniversityofTechnology，Guangzhou510006) Abstract Histopatho】ogica1imagecanrevealthereasonandseverityofdiseases，which isimportant forclinicaldiagnosis．Automaticanalysisofhistopathologicalimagemayreleasedoctor’Sburden for manualannotationwhich can preservemoretimefordoctorstofocuson specialanddifficuItcases ． However，the ambiguous relationship between localregions in a histopath0logical image and histopathologicalcharacteristicsmakesitdifficultto constructa computer—aidmode1．An automatic annotationmethod forhistopathologicalimagesbasedonmultiple—instancemultiple—label(M IM L) learning isproposed，aimingatdirectlymodeling themedicalexperienceofdoctors，whichsuggests thateachannotatedterm associatedwithan imagecorrespondsto alocalvisuallyrecognizedregion． W eproposeaself—adaptiveregioncuttingmethodwithconstraints，tosegmenteachimageintoseveral visuallydisjointregions，andthenperform afeatureextractionforeach generatedregionbasedon textureandinnerstructures．Thewholeimageisregardedasabagandregionsasinstances，thusan imageisexpressed asamultiple—instance sample．Then weproposea sparse ensemblemultiple— instancemultiple—labellearningalgorithm，S—MIM LGP，basedonBayesian learning，and compareit with currentmultiple—instance single labe1and multiple—instance multiple—labelalgorithms． The evaluationonaclinicaldatasetfrom thedermatologyofalargelocalhospitalshowsthattheproposed methodcanyieldmedicallyacceptablean