Transformer-Based Deep Learning for Population-Scale Retinal Image Screening of Ophthalmic Disorders

To perform screening of the retina on a population scale, an automated procedure is required that incorporates accurate, reproducible, interpretable, and computationally costeffective models. Existing approaches using convolutional or transformer architectures typically do not adequately represent both fine-grained pathology and large-scale retinal context simultaneously, which could adversely affect their reliability if used for large-scale applications in clinical practice. In this paper, we propose a hierarchical transformer-based screening framework for retinal fundus images that incorporates patch-based tokenization, global transformer encoding, and hierarchical aggregation of contextual information. We also developed a lightweight prediction head that supports screening for both single and multiple diseases. The framework has been evaluated using standard screening metrics, robustness, and cross-dataset generalization analyses on two eye retinopathy image databases: EyePACS and RFMiD. With regard to screening for a binary outcome of diabetic retinopathy, our method provided an accuracy of 89.4% and an area under the receiver operating characteristic (AUROC) curve of 93.6% on EyePACS and attained an accuracy of 95.2% and a macro-averaged F1 score of 82.7% on RFMiD. Our hierarchical transformer achieved improved robustness to degraded images and increased generalizability across datasets compared with all current state-of-the-art models. The proposed hierarchical transformer demonstrates strong potential for large-scale retinal screening and provides a promising foundation for future clinically validated deployment.