Cluster-filter-based pseudo-label refinement for source-free domain adaptation fundus image segmentation

Source-Free Domain Adaptation (SFDA) enables knowledge transfer without accessing source domain data, effectively enhancing data privacy and security. Consequently, it has been widely applied to optic disc and cup segmentation in fundus images. While SFDA has shown promising results, two key challenges remain: (1) the lack of source domain increases the risk of unstable model adaptation, meaning the reliability of pseudo-labels significantly affects model performance; (2) the low contrast between optic disc and cup, along with relatively small anatomical proportions compared to the background, greatly increases the difficulty of cross-domain segmentation. To address these challenges, this study proposes a plug-and-play module, named Cluster-Filter-based Pseudo-label Refinement (CFPR). Specifically, we exploit the intrinsic characteristics of the data to cluster intermediate model features and obtain clustering filters. These filters are then used to refine the pseudo-labels generated by the teacher model, providing more reliable supervision for student model training and mitigating the performance degradation caused by unstable pseudo-labels. Additionally, to overcome the limitation of the conventional Dice loss function, which struggles to differentiate the importance of the optic disc and cup effectively, we propose a CPG-weighted Dice loss. Based on the class priors guide, this loss function places greater emphasis on hard-to-segment samples, alleviating the effects of class imbalance and the low contrast between the optic disc and cup on model performance. We integrate CFPR with several benchmark models, and experimental evaluations on multiple benchmark datasets demonstrate that our method outperforms existing source-free approaches, particularly in handling ambiguous cup/disc boundaries.