Deep Learning-Based Automated Grading of Diabetic Retinopathy Using EfficientNet-B4 with Attention Mechanisms and Ensemble Fusion

Diabetic retinopathy (DR) is the leading cause of preventable blindness among working-age adults worldwide, with India alone estimated to have over 11.9 million individuals with vision-threatening DR as of 2023. Early detection through systematic fundus screening can prevent up to 95% of severe vision loss; however, the global shortage of trained ophthalmologists — particularly acute in rural and semi-urban India — creates a critical bottleneck in DR screening programmes. Artificial intelligence-based automated DR grading presents a compelling solution, capable of screening large volumes of fundus images accurately and consistently without specialist involvement.This study develops and evaluates a deep learning ensemble system for automated 5-class DR grading (No DR, Mild, Moderate, Severe, Proliferative DR) using EfficientNet-B4 as the primary backbone with dual attention mechanisms — Squeeze-and-Excitation (SE) networks for channel attention and Convolutional Block Attention Module (CBAM) for spatial attention. The ensemble integrates EfficientNet-B4, ResNet-50, and DenseNet-121 predictions through learned weighted averaging. Training and evaluation employ three publicly available benchmark datasets: APTOS 2019 Blindness Detection, IDRiD, and Messidor-2, comprising 7,842 fundus images. Focal loss (γ=2) and SMOTE oversampling address class imbalance. The proposed ensemble achieves macro-average AUC of 0.983, quadratic-weighted kappa of 0.921, sensitivity of 93.6%, and specificity of 96.8% across five DR grades. Gradient-weighted Class Activation Mapping (Grad-CAM) visualisations confirm that the model correctly attends to clinically relevant pathological features including microaneurysms, haemorrhages, hard exudates, and neovascularisation, providing ophthalmologist-interpretable explainability.