Source‐Free Domain Adaptive Fundus Image Segmentation With Multiscale Feature Fusion and Stepwise Attention Integration

ABSTRACT Traditional unsupervised domain adaptation methods usually depend on source domain data distribution for cross‐domain alignment. However, direct access to source data is often restricted due to privacy concerns and intellectual property rights. Without using source data, Source‐free unsupervised domain adaptation methods can align the pre‐trained model with the target domain by generating pseudo‐labels for target domain data, which are then used as labeled samples to guide transfer learning. However, methods that generate pseudo‐labels solely through iterative averaging often neglect spatial correlations among pixels and are susceptible to noise, resulting in blurred label boundaries. To this end, we propose a source‐free domain adaptation framework for fundus image segmentation, which consists of a Multiscale Feature Fusion module for generating high‐quality pseudo‐labels and a Stepwise Attention Integration module for enhancing model training. The Multiscale Feature Fusion module refines the initial pseudo‐labels from the pre‐trained model through neighborhood value filling, effectively reducing noise and sharpening label boundaries. The Stepwise Attention Integration module progressively integrates high‐level and low‐level feature information into the low‐level representation. The fused features preserve high‐resolution details and enrich semantic content, thereby substantially enhancing the model's recognition capability. Experimental results demonstrate that, without using any source domain images or modifying the pre‐trained model, our method achieves performance comparable to or even surpassing state‐of‐the‐art approaches.