Swin-DRNet: A robust transformer framework for diabetic retinopathy screening under heterogeneous imaging conditions

Diabetic retinopathy (DR) has been known as one of the leading preventable causes of vision impairment globally and requires automated screening systems that are reliable to heterogeneous imaging conditions and severely imbalanced classes. While recent deep learning techniques have proven their capabilities within datasets, their limitations regarding large-scale remote screening of ophthalmology are encountered because of their poor generalization from one dataset to another and low sensitivity to minority disease stages. In this publication, we introduce Swin-DRNet, a robust transformer-based framework for automated screening of DR that incorporates multi-stage contrast-adaptive preprocessing, class-aware representation learning, and focal-optimized loss design, all built into the hierarchical architecture of Swin Transformers. This method is optimally suited for improving visibility of lesions, stabilizing the learning process across imbalanced classes, and increasing the robustness of screening performance when using multiple sources of fundus photographs as input. Results from extensive testing of this new approach on three publicly available benchmark datasets APTOS 2019, IDRiD, and Messidor2 demonstrate that the Swin-DRNet achieves consistently high-quality results across all datasets, with an overall accuracy of 96.68%, an F1-score of 96.20%, and a ROC-AUC score of 99.83% on the combined dataset, while also maintaining high recall rates for clinically important advanced DR stages. These findings suggest that the Swin-DRNet provides a reliable and scalable solution for providing diabetic retinopathy screening in the real world, using tele ophthalmology practices that include heterogeneous imaging platforms. To ensure transparency and reproducibility, the complete implementation of the proposed Swin-DRNet framework is publicly available at: https://github.com/damomtpcse/Swin-DRNe.