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

Automated diabetic retinopathy (DR) grading systems often prioritize predictive accuracy over clinical interpretability. While deep learning models can identify disease severity, they frequently fail to explain the relationship between visible lesions, vascular morphology, and clinical biomarkers. This paper addresses this gap by proposing a dual-edge spatial–Jacobian image graph that links image-level predictions to specific, interpretable retinal features.

The authors represent each fundus image as a graph node supported by four aligned evidence streams: vessel information, lesion evidence maps, lesion-based contrastive image embeddings, and morphometric biomarkers (e.g., vessel density, tortuosity). The model constructs two distinct edge families:

• Spatial Branch: Encodes the geometric relationship between lesion evidence and retinal vessels (e.g., whether lesions overlap with arteries or veins).
• Jacobian Branch: Models the sensitivity of lesion-based embeddings to quantitative vascular biomarkers.

A lightweight two-token attention module fuses these branches into a final representation used for five-class DR grading and binary referable-DR classification.

Tested on 2,910 images from the APTOS dataset, the full graph achieved an accuracy of 0.8076 and a quadratic weighted kappa of 0.8312 for five-class grading. For referable DR, the model reached 0.9055 accuracy and 0.9711 AUROC. Beyond performance metrics, the graph structure allows for statistical analysis of lesion-biomarker associations, such as linking neovascularization and hemorrhage evidence to specific vascular caliber and density measures. The authors emphasize that these associations are intended for hypothesis generation rather than causal inference.

This work moves beyond "black-box" classification by decomposing retinal evidence into clinically meaningful components. By explicitly modeling the interaction between lesion location and vascular biomarkers, the framework provides a structured way to investigate how retinal changes correlate with disease progression, offering a potential foundation for more transparent, oculomics-oriented clinical research.