Comparative Evaluation of Deep Ensemble Models for Multi-Stage Diabetic Retinopathy Severity Assessment

Abstract

Diabetic retinopathy (DR), a progressive retinal vascular disorder associated with diabetes mellitus, is a leading cause of visual impairment and blindness globally. Accurate, early-stage classification and severity grading of DR are critical for timely intervention and treatment planning. Deep learning has shown immense promise in automating DR diagnosis, yet the performance of individual models often varies across datasets and disease stages. This study presents a comparative evaluation of deep ensemble learning strategies to enhance the robustness and accuracy of multi-stage DR severity classification. We systematically examine various ensemble methods combining state-of-the-art convolutional neural networks (CNNs) and transformer-based architectures. The analysis incorporates soft voting, weighted averaging, and stacking ensembles applied on benchmark datasets such as APTOS and EyePACS. Evaluation metrics including accuracy, Cohen’s kappa score, sensitivity, and specificity are used to assess model performance. Results demonstrate that ensemble models significantly outperform single-model baselines, especially in differentiating between adjacent DR grades. The proposed ensemble framework offers a promising tool for clinical decision support systems, improving generalizability and reliability in DR detection and grading