Dual-Curriculum and Orthogonal Domain Prototype Learning for Domain Generalization Fundus Image Segmentation

Image segmentation plays a crucial role in retinal disease analysis and computer-aided clinical diagnosis. However, substantial appearance variations across imaging devices, clinical centers, and acquisition conditions often lead to significant performance degradation when segmentation models are deployed to previously unseen domains due to the distribution gap caused by cross-device discrepancies.

A dual-curriculum learning strategy is proposed to handle diverse patterns and mitigate the domain shift between multi-source domain and target domain. Furthermore, a cross-level correlation consistency feature fusion strategy is proposed. This strategy captures the pixel-level correlations from low-level feature and maintains these correlations within high-level features. In addition, orthogonal domain prototypes of multi-source domains are constructed to better adapt to the multi-source domain generalization task.

Experiments on two multi-domain OCT image datasets demonstrate that the proposed method consistently outperforms state-of-the-art approaches. The results validate its effectiveness under cross-domain scenarios. Additional experiments on a color photography image dataset further confirm the robustness and generalizability of the proposed framework across diverse fundus imaging modalities.

Our model trained on labeled fundus images from various manufacturers' devices achieves superior domain generalization when tested on fundus images from previously unseen devices.

The proposed method is suitable for routine clinical use. By mitigating the need for device-specific annotations, the proposed method holds promise for routine clinical deployment and reducing manual labeling costs.