Source-Free Domain Adaptation Fundus Image Segmentation Based on Semantic-Aware Adversarial Learning

In recent years, fundus image segmentation has become a fundamental task in computer-aided diagnosis of ophthalmic diseases. However, the performance of segmentation models severely degrades when they are transferred across domains, primarily due to domain shift and the lack of reliable annotations in the target domain. Source-Free Domain Adaptation (SFDA) provides a feasible solution by adapting a pre-trained source model to the target domain without requiring access to source data. Nevertheless, the presence of noisy pseudo-labels and the absence of structural alignment remain challenging issues that limit the effectiveness of existing methods. To address these problems, this paper proposes a Semantic-Aware Adversarial Learning (SAAL) framework for source-free domain adaptation in fundus image segmentation, which is designed with two main components. First, a triple pseudo-label filtering mechanism is introduced, integrating confidence estimation, uncertainty evaluation, and class prototype consistency to ensure high-quality supervision. Second, a dual-branch discriminator is developed, which performs both domain discrimination and semantic classification, achieving pixel-level semantic alignment while maintaining domain-invariant representations. This design mitigates the impact of noisy labels and enhances structural consistency across domains. Experimental evaluations on multiple benchmark fundus image datasets demonstrate that the proposed method outperforms existing SFDA approaches, particularly in handling ambiguous cup/disc boundaries.