南京师范大学学报（工程技术版）2025，Vol.25Issue(2)：79-87,9.DOI:10.3969/j.issn.1672-1292.2025.02.007
基于无监督域自适应和Transformer的视网膜图像语义分割
Unsupervised Domain Adaptation and Transformer-based Retinal Image Semantic Segmentation
摘要
Abstract
To Address the challenges of semantic segmentation in retinal images with complex backgrounds,significant vascular structure variations,and edge information loss,this paper proposes an improved Transformer-based unsupervised domain adaptation(TUDA)method.Firstly,the model consists of an enhanced Transformer encoder and a context-aware feature fusion decoder.The encoder effectively handles complex interactions between different positions in retinal images by fusing local information.Secondly,to reduce the domain gap,an intermediate domain is introduced between the source and target domains,and a dual-teacher network is used for alternating training.Finally,high-resolution features are generated by using HRNet to preserve more edge information in retinal images.Compared to conventional medical image semantic segmentation methods,the proposed method achieves intersection-over-union(IoU)of 65.36%and 69.79%,and sensitivity(SEN)of 82.51%and 85.56%on CHASE_DB1 and DRIVE datasets,respectively,demonstrating superior segmentation performance.关键词
语义分割/中间域/无监督域自适应/医学图像Key words
semantic segmentation/intermediate domain/unsupervised domain adaptation/medical images分类
计算机与自动化引用本文复制引用
孙成富,李敏,邹佳辰,秦思奇,季嘉杰,孔维龙..基于无监督域自适应和Transformer的视网膜图像语义分割[J].南京师范大学学报（工程技术版）,2025,25(2):79-87,9.基金项目
国家自然科学基金青年科学基金项目(6200050273). （6200050273）
