Diabetic retinopathy (DR) is among the leading causes of preventable blindness worldwide. Convolutional neural networks process fundus images in isolation, overlooking the inter-patient structural regularities that are clinically informative, especially at inter mediate disease stages. We propose a topology-aware, graph-based deep learning framework that enriches CNN embeddings with persistent homology descriptors of retinal vessel net works and leverages a population-level similarity graph re ned by GraphSAGE. Evaluated on Kaggle DR, Messidor-2, and APTOS 2019, the proposed method achieves accuracies of 95.5%, 96.1%, and 94.6%, respectively, outperforming the CNN baseline by 1.52.3 percent age points and improving QWK from 0.84 to 0.88. Ablation studies con rm the comple mentarity of topological and relational components, and statistical analysis validates that vascular topological complexity is a signi cant discriminator of DR severity.
Belhadj et al. (Thu,) studied this question.