Federated Learning for Diabetic Retinopathy Detection Using Vision Transformers

A common consequence of diabetes mellitus called diabetic retinopathy (DR) results in lesions on the retina that impair vision. It can cause blindness if not detected in time. Unfortunately, DR cannot be reversed, and treatment simply keeps eyesight intact. The risk of vision loss can be considerably decreased with early detection and treatment of DR. Ophtalmologists must manually diagnose DR retinal fundus images, which takes time, effort, and is cost-consuming. It is also more prone to error than computer-aided diagnosis methods. Deep learning has recently become one of the methods used most frequently to improve performance in a variety of fields, including medical image analysis and classification. In this paper, we develop a federated learning approach to detect diabetic retinopathy using four distributed institutions in order to build a robust model. Our federated learning approach is based on Vision Transformer architecture to classify DR and Normal cases. Several performance measures were used such as accuracy, area under the curve (AUC), sensitivity and specificity. The results show an improvement of up to 3% in terms of accuracy with the proposed federated learning technique. The technique also resolving crucial issues like data security, data access rights, and data protection.