Leveraging Graph Neural Networks for Topological Feature Extraction and Severity Classification of Diabetic Retinopathy
Gattu Tejaswini,
Rapolu Harika,
Singireddy Varsha,
Shaik Afzal
Diabetic Retinopathy (DR) is a leading cause of blindness in diabetic patients, requiring accurate and
early severity classification from retinal fundus images. Traditional Convolutional Neural Networks
(CNNs) often overlook complex relational and structural patterns in retinal vasculature and lesions. This
paper proposes a novel approach leveraging Graph Neural Networks (GNNs) for topological feature
extraction and multi-class severity classification of DR (No DR, Mild, Moderate, Severe, Proliferative).
Retinal images are preprocessed and transformed into graph representations where nodes capture local
features (e.g., lesions, vessels) and edges model topological relationships. A Variational Autoencoder
(VAE) extracts latent embeddings, followed by a Graph Convolutional Neural Network (GCNN) that
aggregates neighborhood information to learn discriminative topological patterns. The model is
evaluated on the EyePACS and APTOS datasets, demonstrating superior performance in handling class
imbalance and capturing subtle structural changes. Results show improved accuracy, F1-score, and
Cohen’s Kappa compared to baseline CNNs, enhancing reliability for automated DR screening. The
system promotes explainable AI through graph-based interpretability while maintaining computational
efficiency.
