GBLR-Net: Graph-Based Lesion Relationship Network for Diabetic Retinopathy Classification through Segmented Retinal Features

Abstract

Diabetic retinopathy [DR] is the leading cause of vision loss in humans, and disease severity is determined not only by the presence of retinal lesions but also by their spatial distribution and inter-lesion relationships. There is a reliance on deep learning (DL) approaches that use convolutional neural networks (CNNs), which primarily capture local pixel-level patterns, limiting their ability to model clinically meaningful lesion interactions. To mitigate this, the present study incorporates the two-stage graph network that integrates lesion-level segmentation with graph-based representation learning for diabetic retinopathy classification. In the first stage, a U-Net architecture with ResNet50 backbone is employed to perform precise multi-class segmentation of pathological retinal lesions from fundus images. In the second stage, individual lesion instances are extracted from the segmentation maps and represented as nodes in a graph, where node attributes encode morphological and spatial characteristics, and edges capture spatial proximity relationships among lesions. A graph neural network (GNN) is then applied for learning high-level features for disease grading. Experimental evaluations conducted on publicly available retinal datasets demonstrate that the proposed lesion-aware graph-based model achieves improved classification performance compared to conventional convolutional baselines, particularly in distinguishing intermediate disease stages. The results indicate that explicitly modeling lesion relationships enhances both diagnostic accuracy and interpretability, offering a clinically relevant and extensible solution for automated diabetic retinopathy assessment. Our model achieves state-of-the-art performance on the Messidor-2 dataset, outperforming most existing methods across accuracy, precision, recall, and F1 score. 99.08%, 99.40%, 99.20%, and 99.10%, respectively that demonstrating its superior effectiveness compared to prior works.