Abstract—Diabetic Retinopathy (DR) remains one of the leading causes of preventable blindness worldwide, and early, reliable diagnosis is essential for reducing vision loss. Deep learning has shown promise in this domain, but single models often suffer from limited generalizability, sensitivity–specificity imbalance, and high computational demand. To address these challenges, we present PolyVision, a modular ensemble framework designed for robust and equitable DR screening. PolyVision integrates three complementary backbones—ResNet50, EfficientNet-B2, and Vision Transformer—each capturing different levels of spatial and contextual retinal features. Their predictions are combined through a dual fusion mechanism based on mean and maximum voting, which balances diagnostic sensitivity and specificity while minimizing variance across models. To further enhance robustness, the models are trained with diverse augmentation strategies, and hyperparameters are tuned for optimal performance. Evaluated on ultra-widefield fundus images, PolyVision achieved an AUC-ROC of 0.953, an AUPRC of 0.975, and an inference latency of 110 ms per image, demonstrating both high diagnostic accuracy and clinical efficiency. Beyond accuracy, the framework incorporates fairness evaluation across imaging subgroups, supporting more equitable diagnostic outcomes. Its lightweight design also facilitates deployment in resource-constrained clinical settings without compromising reliability. These results highlight the potential of ensemble learning to provide scalable, accurate, and fair DR screening. However, additional validation on multi-institutional datasets and real-world clinical environments remains necessary before broad clinical adoption.
 
Keywords—diabetic retinopathy, convolutional neural networks, vision transformer, deep learning, model fusion, fairness in AI, medical image analysis


Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).