王成前

多源域自适应旨在利用多个不同来源的数据来协同提升模型在无标签目标域上的分割表现，但如何有效融合异构特征并抑制跨域过程中的负迁移效应，仍是当前该领域的难点与挑战。本文提出了一种基于张量分解与斯皮尔曼对齐的多源域自适应分割算法（TS-MDA）。该算法以具备全局感知能力的U-Mamba为骨干网络以捕捉宏观解剖结构表征，并结合张量低秩分解思想在参数空间重构线性投影层；在深层特征处理上，算法构建对比学习以实现跨域语义对齐，并通过分组向量量化精炼深层表征。本文在公开眼底OCT数据集上进行全面的对比实验评估。

OCT图像；多源域自适应；图像分割；张量分解；表征学习

[1] HeJ,JiaX,ChenS,etal.Multi-sourcedomainadaptationwithcollaborativelearningforsemanticsegmentation[C]//ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecognition.2021:11008-11017. [2] MaJ,LiF,WangB.U-mamba:Enhancinglong-rangedependencyforbiomedicalimagesegmentation[C]//InternationalConferenceonMedicalI magingwithDeepLearning.2024. [3] PengX,BaiQ,XiaX,etal.Momentmatchingformulti-sourcedomainadaptation[C]//ProceedingsoftheIEEE/CVFinternationalconferenceonc omputervision.2019:1406-1415. [4] HuS,LiaoZ,XiaY.Domainspecificconvolutionandhighfrequencyreconstructionbasedunsuperviseddomainadaptationformedicalimageseg mentation[C]//InternationalConferenceonMedicalImageComputingandComputer-AssistedIntervention.Cham:SpringerNatureSwitzerland,2022:650-659. [5] PeiC,WuF,YangM,etal.Multi-sourcedomainadaptationformedicalimagesegmentation[J].IEEEtransactionsonmedicalimaging,2023,43(4):1640-1651. [6] KimS,JoSH,HaW,etal.DistributionallyRobustClassificationforMulti-sourceUnsupervisedDomainAdaptation[C]//InternationalConferenceonLearningRepresentations.2026. [7] BogunovićH,VenhuizenF,KlimschaS,etal.RETOUCH:TheretinalOCTfluiddetectionandsegmentationbenchmarkandchallenge[J].IEEEtra nsactionsonmedicalimaging,2019,38(8):1858-1874. [8] LiY,YuanL,ChenY,etal.Dynamictransferformulti-sourcedomainadaptation[C]//ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecognition.2021:10998-11007. [9] TasarO,TarabalkaY,GirosA,etal.StandardGAN:Multi-sourcedomainadaptationforsemanticsegmentationofveryhighresolutionsatelliteima gesbydatastandardization[C]//ProceedingsoftheIEEE/CVFConferenceonComputerVisionandPatternRecognitionWorkshops.2020:192-193.

点击下载PDF