Graph-based representation of retinal lesions for an interpretable diagnosis of diabetic retinopathy

Diabetic retinopathy (DR) is a leading cause of blindness among the working-age population worldwide. DR diagnosis and grading are based on identifying characteristic retinal lesions through fundus imaging. Despite the effectiveness of deep learning techniques in DR detection and grading, these methods often lack interpretability. This work introduces a novel approach that leverages a graph representation of the retina, where each node corresponds to a lesion, and a graph neural network (GNN) is used to grade DR. Our method aligns with clinical guidelines by using lesion-specific information while maintaining the capacity of deep learning models. We first segment DR lesions using a pre-trained convolutional neural network (CNN) and then construct a graph with lesions as nodes connected to their k nearest neighbours. Features for each node are derived from the lesion-specific regions in the fundus image. The resulting lesion graph is classified using a graph attention network (GAT) to make the DR grade prediction. Our method was evaluated on multiple public datasets, achieving performance comparable to state-of-the-art techniques based on quadratic weighted Cohen’s kappa and other metrics. This graph-based approach offers a balance between local lesion segmentation and global image classification, potentially enhancing interpretability and robustness for clinical applications.