Dual-edge graph neural network makes diabetic retinopathy grading explainable

A new research paper on arXiv introduces a graph-based neural network that brings explainability to automated diabetic retinopathy (DR) grading. Traditional deep learning models for fundus image analysis often produce accurate predictions but fail to show how features like lesion distribution around retinal vessels drive the final grade. The proposed dual-edge spatial-Jacobian image graph directly addresses this gap by constructing a rich graph representation from each fundus image.

The graph fuses four aligned evidence streams: AutoMorph-derived vessel information ($X_1$), DR-XAI-style lesion evidence maps ($X_2$), a 128-dimensional lesion-based contrastive image embedding ($X_3$), and AutoMorph morphometric biomarkers ($X_4$). Two edge families encode different relationships: a spatial edge branch ($X_{12}$) captures vessel-lesion geometry, while a Jacobian branch ($X_{34}$) models sensitivity between embeddings and biomarkers. A lightweight two-token attention mechanism fuses both edge families into a final image graph. Tested on 2,910 non-augmented APTOS images, the model achieves 0.8076 accuracy, 0.8312 quadratic weighted kappa, 0.5915 macro-F1, 0.9330 adjacent-grade accuracy, and 0.9055 accuracy with 0.9711 AUROC for referable DR. The authors position the framework as an explainable representation-learning tool for lesion-biomarker hypothesis generation, explicitly noting it is not yet a deployment-ready clinical classifier. Code is available on GitHub.