A Dual Edge Spatial Jacobian Image Graph for Interpretable Diabetic Retinopathy Grading

2026-06-23

AI Summary

This study addresses the limited interpretability of existing automated diabetic retinopathy (DR) grading methods, which often overlook the spatial relationships between lesions and retinal vasculature as well as their associations with biomarkers. To overcome this, the authors propose a bilateral spatial Jacobian image graph model that jointly captures lesion–vessel geometric relationships and embedding–biomarker sensitivity for the first time. The model integrates vascular structure, lesion evidence maps, contrastive embeddings, and morphometric biomarkers, enhanced by a lightweight dual-label attention mechanism for effective multi-source information fusion. Evaluated on the APTOS dataset (2,910 images), the method achieves an accuracy of 0.8076, weighted Kappa of 0.8312, macro F1-score of 0.5915, and adjacent-level accuracy of 0.9330. For referable DR detection, it attains 0.9055 accuracy and 0.9711 AUROC, substantially improving both diagnostic performance and model interpretability.

Abstract

Automated diabetic retinopathy (DR) grading from colour fundus photographs can achieve strong predictive performance, but clinical interpretation requires more than an image-level label. It requires understanding how lesion evidence is distributed around retinal vessels and how this evidence relates to quantitative vascular biomarkers. We present a dual-edge spatial-Jacobian image graph for interpretable DR grading. Each fundus image is represented as a graph node with four aligned evidence streams: AutoMorph 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 ($X_4$). The spatial edge branch ($X_{12}$) encodes vessel-lesion geometry, while the Jacobian branch ($X_{34}$) models embedding-biomarker sensitivity. Lightweight two-token attention fuses both edge families into a final image graph. On 2,910 matched non-augmented APTOS images, the full graph achieves 0.8076 accuracy, 0.8312 quadratic weighted kappa, 0.5915 macro-F1, and 0.9330 adjacent-grade accuracy; referable DR reaches 0.9055 accuracy and 0.9711 AUROC. The framework is positioned as an explainable representation-learning tool for lesion-biomarker hypothesis generation, rather than as a deployment-ready clinical classifier. The code is available at https://github.com/Inamullah-Colab/dual-edge-dr-graph-xai.

Problem

• diabetic retinopathy
• interpretable grading
• lesion evidence
• vascular biomarkers
• explainable AI

Innovation

• dual-edge graph
• spatial-Jacobian modeling
• interpretable DR grading
• lesion-biomarker correlation
• contrastive image embedding