Ensemble of graph neural networks with Transformer fusion for diabetic retinopathy severity classification

Diabetic retinopathy (DR) is a major cause of vision impairment among working-age diabetics, but its early stages usually progress without noticeable symptoms. Manual modes of screening, in contrast to machine techniques, are labor-intensive, subjective, and prone to wide inter-observer variability, leading to a dire need for such automated, reliable, and scalable diagnostic tools. Moreover, DR datasets often suffer from severe class imbalance, where early and advanced stages are underrepresented, leading to limited model robustness and reduced accuracy across all severity levels. To tackle the challenges posed by these limitations, we propose a new graph-based deep learning framework modeling retinal vasculature as graphs to capture clinically relevant topological features. The proposed method relies on a heterogeneous ensemble of graph neural networks (GNNs)-graph attention networks (GATs), graph convolutional networks (GCNs), and GraphSAGEs-each analyzing different structural patterns. Their outputs get adaptively fused via a Transformer encoder to allow for contextually informed feature aggregation. To confront the class imbalance, the framework relies upon a self-attention conditional generative adversarial network (SA-CGAN) to generate high-quality synthetic fundus images for the underrepresented DR stages, thus boosting model robustness and sensitivity. Multiple experiments on the APTOS 2019 dataset demonstrate the promising nature of the framework, achieving an accuracy of 92.4%, an F1-score of 92.7%, and an AUC-ROC of 0.96, which clearly states that it outperforms all CNN-based benchmark models and single-GNN models. This performance validates the use of heterogeneous GNN reasoning in combination with adaptive fusion through a Transformer and generative augmentation of underrepresented data; hence, creating a clinically interpretable and scalable automated solution for DR screening in all severity stages.