Abstract
Diabetic retinopathy (DR) is a leading cause of vision impairment, which significantly impacts working-class populations, necessitating accurate and early diagnosis for effective treatment. Traditional DR classification relies on Convolutional Neural Network (CNN)-based models and extensive preprocessing. In this work, we propose a novel approach leveraging pre-trained models for feature extraction, followed by Graph Convolutional Networks (GCNs) for refined embedding representation. The extracted feature vectors are structured as a graph, where GCN enhances embeddings before classification. The proposed model incorporates quality assessment by predicting a confidence score through a dedicated fully connected layer, trained to align with ground truth quality using binary cross-entropy loss. Uncertainty estimation is achieved by calculating the variance across multiple stochastic passes, providing a measure of the model's prediction reliability. We evaluate the proposed DR detection approach on APTOS2019, Messidor-2, and EyePACS datasets, achieving superior performance over state-of-the-art methods. Using MobileViT as the main feature extractor, we reached a remarkable 98.45% accuracy, 98.45% F1-Score, and 98.06% Kappa on the APTOS2019 dataset. The DenseNet-169 proved to be the best backbone of the pipeline for the Messidor-2 dataset, with an accuracy of 94.90%, F1-Score of 94.87%, and Kappa of 93.63%. Additionally, for external validation, the model demonstrated strong generalization capability on the EyePACS dataset, where DenseNet-169 achieved 97.38% accuracy, 97.37% F1-Score, and 96.72% Kappa, while MobileViT obtained 96.02% accuracy, 96.02% F1-Score, and 95.03% Kappa. Our innovative architecture incorporates uncertainty estimation and quality assessment techniques, enabling accurate confidence scores and enhancing the model's reliability in clinical environments. Furthermore, to strengthen interpretability and facilitate clinical validation, Grad-CAM heatmaps were employed to demonstrate the significance of different input regions on the model's predictions.