An Interpretable Deep Learning Framework for Diabetic Retinopathy Detection Using Patch-Level Variational Autoencoders and Graph Attention Networks

Abstract

Diabetic Retinopathy (DR) is a sight-threatening complication of diabetes that affects millions worldwide. Early detection and accurate grading of DR severity are critical for timely treatment. In this work, we propose an interpretable and modular pipeline that combines patch-level Variational Autoencoders (VAE) for compact feature encoding, spatial graph construction to capture anatomical context, and Graph Attention Networks (GAT) for classification with region-level interpretability. The approach emphasizes clinical relevance by producing attention heatmaps that align with lesion locations. We evaluate the method on the EyePACS dataset and report competitive performance across five DR grades while offering improved interpretability over conventional end-to-end CNNs.