Graph hybrid neural network integration for enhanced diabetic retinopathy detection in retinal image

Abstract

Diabetic retinopathy (DR) is a progressive microvascular complication of diabetes that affects the retina. This study presents a hybrid graph neural network (GNN) architecture that combines a graph isomorphism network (GIN) with GraphSAGE to improve performance and interpretability in diabetic retinopathy (DR) classification. The framework captures both local and broader lesion relationships, which enhances classification–especially when labeled data are scarce. GIN achieves higher recall and F1 scores, while GraphSAGE yields better precision and ROC-AUC; either method can be chosen depending on diagnostic priorities. Overall, these findings show that GNN-based models can strengthen DR diagnosis, support earlier detection, and help reduce ophthalmologists’ workload by effectively modeling retinal image structure.