Learning Generalizable Diabetic Retinopathy Grading by Decoupled State Space Decoding

Diabetic retinopathy (DR) is a leading cause of vision loss. The rapid advancement of deep learning has significantly propelled the development of automated DR grading methods. However, retinal images are typically collected from diverse clinical centers using equipments from various vendors, leading to domain shifts. As a result, the pre-trained model for DR diagnosis has to handle the unseen retinal images when doing inference. In this paper, we propose Decoupled State Space Decoding (DSSD), a selective state space-based model designed to enhance the generalization capability of DR grading methods to unseen domains. Leveraging the selective scan mechanism to encode local patches relevant to lesions and the recurrent modeling to capture long-range dependencies, DSSD aims to mitigate cross-domain style variations through both channel-wise and sample-wise decoupling. Additionally, it introduces an implicit constraint between shallower and deeper state embeddings to ensure stable recurrent modeling for grade-level predictions. Experiments conducted under both leave-one-domain-out and extreme-single-domain-generalization settings demonstrate its state-of-the-art performance.