Unsupervised Hybrid-Rectification Domain Adaptation Model for Source-Relaxed Fundus Image Segmentation

Glaucoma is a common and clinically significant chronic eye disease that can lead to irreversible vision loss if not treated in time, highlighting the urgent need for efficient automatic segmentation methods to assist in diagnosis and treatment. Although various approaches have attempted to improve the performance of target models via pseudo-label generation, self-supervised learning, or model uncertainty estimation, three major challenges remain. First, source models are often insensitive to fine-grained features. Second, pseudo-labels generated from target domain images with blurry structures are extremely sparse and nearly ineffective for training, resulting in negative transfer. Third, the training of the target model relies entirely on unreliable supervision, which easily leads to error accumulation. To address these challenges, we introduce a Dual-domain Collaborative Augmentation (DCA) mechanism for the first time, which enhances the model's ability to perceive critical ocular details by applying noise perturbation in the frequency domain and data augmentation in the spatial domain. We also propose an uncertainty-guided hybrid refinement strategy to mitigate the adverse effects of domain discrepancy, along with a novel confidence-aware loss function to optimize the training process of the target model. The effectiveness of our method was experimentally validated on three fundus datasets under two domain adaptation settings, demonstrating its great potential in early glaucoma screening.