Deep Convolutional Neural Network Architectures for Automated Diabetic Retinopathy Grading from Fundus Retinal Images: A Comparative Performance Analysis

Diabetic retinopathy (DR) is the leading cause of preventable blindness in working-age adults globally, affecting an estimated 93 million people worldwide, with India bearing the second-largest absolute burden due to its 77-million-strong diabetic population. Early-stage DR is asymptomatic, making systematic screening of the diabetic population essential for timely intervention. This paper presents a comparative evaluation of four deep learning architectures — a custom 7-layer CNN, VGG-16, ResNet-50, and EfficientNet-B0 — for automated five-class DR grading (No DR, Mild, Moderate, Severe, Proliferative DR) on a 10,000-image fundus dataset curated from three district hospital ophthalmology units in Lucknow and Raipur. Transfer learning with ImageNet pre-training, class-weighted focal loss to address the 8:1 class imbalance between No DR and Proliferative DR, and extensive data augmentation (rotation, horizontal flip, CLAHE contrast enhancement) were applied uniformly across architectures for fair comparison. ResNet-50 achieved the highest test accuracy of 96.8% and macro-averaged F1-score of 95.9%, with AUC of 0.97 on the ROC curve. EfficientNet-B0 showed competitive performance (accuracy 95.4%) with 40% fewer parameters. A confusion matrix analysis reveals that the most clinically significant misclassification — Severe DR predicted as Moderate (false negative rate 8.1%) — warrants human-in-the-loop verification for grades 2–3. The study demonstrates the feasibility of automated DR screening deployment on hospital-grade computing infrastructure in resource-limited settings.