DR-MAE: Self-supervised learning for diabetic retinopathy grading based on masked autoencoder

Precise grading of diabetic retinopathy (DR) is crucial for preventing vision loss. Existing deep learning methods face challenges including scarcity of annotated data, difficulty in identifying subtle lesions, and insufficient local–global feature modeling. This paper proposes DR-MAE, an anatomy-guided masked autoencoder self-supervised learning framework. The main innovations include: (1) DR-Masking strategy, which generates intelligent masks based on optic disc localization and vascular topology analysis, avoiding the destruction of critical regions by random masking; (2) Hybrid Dimensional Convolution Block (HDC Block), which integrates multi-scale dynamic modules and global context attention modules to achieve multi-scale feature extraction and local–global feature interaction. Experiments on APTOS 2019 and DDR datasets demonstrate that DR-MAE achieves grading accuracies of 91.30% and 86.64% respectively, representing state-of-the-art performance. Ablation experiments validate the effectiveness of each component, proving the advantages of anatomy-guided self-supervised learning in DR grading.