A graph-based deep learning framework for diabetic retinopathy classification with topology-aware feature augmentation

Abstract

Diabetic retinopathy (DR) is the leading cause of preventable blindness in working-age adults, affecting an estimated 103 million people worldwide. Standard deep learning classifiers treat fundus images as independent samples, ignoring latent inter-patient relational structure that is most informative at clinically ambiguous intermediate severity levels. We propose a topology-aware, graph-based deep learning framework combining three complementary components: (i) an EfficientNet-B3 convolutional backbone for high-level visual feature extraction; (ii) persistent homology descriptors (H0 and H1) derived from morphologically skeletonised retinal vascular networks, characterising global vascular topology in a noise-robust manner; and (iii) a GraphSAGE graph neural network propagating disease-related information across a population-level similarity graph, refining representations through inductive neighbourhood aggregation. The similarity graph combines cosine similarity on visual features with 2-Wasserstein distance between persistence diagrams. Evaluated on three public benchmarks, the framework achieves 95.5% accuracy on Kaggle DR, 96.1% on Messidor-2, and 94.6% on APTOS 2019, consistently outperforming a strong CNN baseline by 1.5-2.3 percentage points across accuracy, Quadratic Weighted Kappa, and macro-F1. Ablation experiments confirm synergistic contributions of topological feature augmentation and relational graph learning. One-way ANOVA (F > 80, p < 0.001) confirms that DR progression is reflected in global vascular topology across all five severity stages, providing quantitative biological grounding for the framework design. Code and data are publicly available at https://github.com/Nader-BelHadj/plosene.