Automated Diabetic Retinopathy Severity Classification using EfficientNet-B3 with Grad-CAM Explainability

Diabetic Retinopathy (DR) is a progressive microvascular complication of diabetes and one of the leading causes of preventable blindness worldwide. With the global rise in diabetes prevalence, millions of individuals are at risk of vision impairment due to delayed diagnosis and inadequate screening infrastructure. Early and accurate identification of DR severity is therefore essential to reduce irreversible retinal damage and improve patient outcomes. However, manual assessment of retinal fundus images is time-consuming, resource-intensive, and dependent on specialist expertise, creating an urgent need for reliable automated screening systems. This study presents an EfficientNet-B3–based deep learning framework for multi-stage DR severity classification using a clinically annotated retinal fundus dataset. The proposed system integrates standardized preprocessing, structured fine-tuning, and agreement-based evaluation to enhance diagnostic reliability. Experimental results demonstrate strong performance, with class-wise accuracy ranging from 0.945 to 0.981, and Area Under the Curve (AUC) values exceeding 0.948 across all categories. Agreement-based evaluation further confirms robustness, achieving Cohen’s Kappa (0.932) and Quadratic Weighted Kappa (0.961), indicating high ordinal consistency. The overall test accuracy derived from the confusion matrix is approximately 91.0%, with minimal misclassification between adjacent severity levels. These results validate the effectiveness, scalability, and clinical reliability of the proposed framework for automated DR screening and severity grading, supporting improved early detection and reduced global burden of diabetes-related blindness.