融合U-Net和Transformer的医学图像分割算法

打开文本图片集

中图分类号：TP391 文献标识码：A

文章编号：2096-4706（2025）19-0062-08

Abstract：To addressthe limitations ofU-Netnetworks incapturing global features，which restricts their performance inmedicalimagesegmentation whereshapeand structuralvariationsaresignificantandtheissueofinsuffcientlocalization capability in Transformer networks when used independentlyduetolackofsuffcientlocalfeatures，this paper proposesadeep segmentationframeworkcalled SeriTansUNet.Firstly，theTransformermoduleisembeddedinthebottleneckpartofU-Netto enhance globalcontextualawarenes.Secondly，aFeatureEnhancementBlock（FEB）isintroducedbetweentheConvolutional NeuralNetwork（CNN）encoderofU-NetandtheTransformer toenrich feature informationfrommulti-semantic spacesand multi-dimensionalchanels.FinallaFusionmoduleisdesignedintheskioectionstodeelyintegratefeatureseractedby theTransformerand CNNencoderalong thechanel dimension，fullyleveragingthestrengthsofboth toimprovesegmentation accuracy.ExperimentsdemonstratethattheproposedSeriTransUNetframeworkachievesoutstandingsegmentationperformance on the Synapse dataset，with aDSC of 81.53% and a HD of 24.15.

Keywords：CNN;Transformer;Feature EnhancementBlock（FEB）;Fusionmodule

0 引言

医学影像智能分析领域中，腹部CT图像的多器官分割技术具有重要的临床应用价值。（剩余12935字）