基于改进TransUNet的肺部图像分割

打开文本图片集

中图分类号：TN911.73-34；TP391 文献标识码：A 文章编号：1004-373X（2025）15-0027-10

Lung image segmentation based on improved TransUNet

SHI Yongtaol，²，QIU Kangqi，²，LIU Di1，2 ，DUWei1，² （1.CollegeofComputerand Information Technology，China ThreeGorges UniversityYichang 443oo2，China; 2.HubeiKeybelVisodi

Abstract：Semanticsegmentation，asacrucial stepin lung imageanalysis，directlyafects theaccuracyoffurther image analysisandtreatmentdecisions.Thelimitationsofthetraditionalsegmentationmethods inthefaceof rregularshapes，blurrd boundaries，andnoiseoflungorgansarelowauracyandsusceptibilitytoerrorsinboundarysegmentation.Inviewof this，a multi-scaleedgefeaturefusionbasedneural network （MSB-AfTransU2Net）isproposedforlung imagesegmentation.Firstlythe encoder and decoder in TransUNet are replaced，and the RSU module of U2 -Net is used to enhance the performance of feature extraction.Then，theatentionfeature fusionmechanism isusedtoreplacetheoriginalConcatmethod，soastoreducemodel parametersandimprovethefusionefectoffeatures.Next，amulti-scalefeatureextractorandaboundary-guidedcontext aggregation moduleareadded tofuse and extract more accuratelung edgefeatures.Finally，Dicelossand crossentropylossare adoptedtoreateanovellossfunction inorder tooptimizethemodellossfunction.Theefectivenessof theproposedalgorihm wasvalidatedontheCOVID-19RadiographyDatabase dataset experimentally.The experimental resultsshow thattheMSBAffTransU2Net has improvedthe prospect intersection union ratio （pIoU）and mean acuracy （mAcc）on the COVID dataset by （204 3.03% and 0.72% ，respectively，in comparison with the TransUNet algorithm.

Keywords：COVID-19;lung image segmentation; TransUNet; edge feature;boundary-guided context aggegation module; attentional feature fusion

0 引言

自2019年12月以来，新型冠状病毒（C0VID-19）疾病给全世界人类的健康带来了前所未有的挑战[-2]。（剩余13079字）