Automatic classification of diabetic retinopathy using convolutional neural networks: A deep learning-based ensemble approach

Diabetic retinopathy (DR) is one of the leading causes of preventable blindness. This study presentsan integrated approach for automatic classification of DR stages from retinal fundus images, using twelve convo-lutional neural network (CNN) architectures and an efficient ensemble (EfficientNet-B3 + VGG-19). A total of5,842 images were processed through a pipeline including CLAHE, Gaussian filtering, and normalization, alongwith data augmentation to address class imbalance. The ensemble outperformed individual models, achieving aglobal accuracy of 80.1% and a quadratic weighted Kappa of 0.78. Sensitivity was robust for advanced stages (upto 87.3%), but remained low for mild DR (21.0%), highlighting the ongoing challenge in early-stage detection.