文献速递基于深度学习的超声图像前列腺
文章题目:Accurateandrobustdeeplearning-basedsegmentationoftheprostateclinicaltargetvolumeinultrasoundimages
研究人员:DavoodKarimietal.
研究单位:DepartmentofElectricalandComputerEngineering,UniversityofBritishColumbia,Vancouver,BC,Canada
期刊名称:MedicalImageAnalysis
影响因子与分区:11.(Q2)
01核心亮点
ComparedtoothertraditionalandCNN-basedmethods,ourmethodachievedsignificantlybetterresultsintermsofHD,whichmeasureslargestsegmentationerror.(与其他传统的和基于CNN的方法相比,我们的方法在HD方面取得了更好的效果,HD是衡量最大分割误差的指标。)
Moreover,ourmethodalsosubstantiallyreducedthemaximumerrorsonthepopulationoftestimages.(此外,我们的方法还大大减少了测试图像总体上的最大误差。)
02
思路与方法
基于超声图像建立一种在近距离放射治疗的经直肠超声(TRUS)图像中对前列腺临床靶体积进行准确而稳健的自动分割的方法,并测试该方法的分割精度。
03
摘要
Thegoalofthisworkwastodevelopamethodforaccurateandrobustautomaticsegmentationoftheprostateclinicaltargetvolumeintransrectalultrasound(TRUS)imagesforbrachytherapy.Theseimagescanbedifficulttosegmentbecauseofweakorinsufficientlandmarksorstrongartifacts.Wedeviseamethod,basedonconvolutionalneuralnetworks(CNNs),thatproducesaccuratesegmentationsoneasyanddifficultimagesalike.Weproposetwostrategiestoachieveimprovedsegmentationaccuracyondifficultimages.First,forCNNtrainingweadoptanadaptivesamplingstrategy,wherebythetrainingprocessisencouragedtopaymoreattentiontoimagesthataredifficulttosegment.Secondly,wetrainaCNNensembleandusethedisagreementamongthisensembletoidentifyuncertainsegmentationsandtoestimateasegmentationuncertaintymap.Weimproveuncertainsegmentationsbyutilizingthepriorshapeinformationintheformofastatisticalshapemodel.OurmethodachievesHausdorffdistanceof2.7±2.3mmandDicescoreof93.9±3.5%.Comparisonswithseveral
转载请注明:http://www.shijichaoguyj.com/wxjs/9131.html