Diabetic Retinopathy Detection and Severity Classification: A Vision Transformer Model Approach and Federated Learning Approach

Diabetic Retinopathy is a severe complication of diabetes, create a significant threat to vision, leading to irreversible blindness if it is not detected in the early stages. This study explores the integration of Deep Learning, a powerful subset of Machine Learning, for the accurate and effective detection of DR. Additionally, privacy concerns are addressed through the utilization of a decentralized approach known as Federated Learning. The Vision Transformer Model, Federated Averaging Model, and Federated Proximal Model. These models collectively aim to classify retinal images as either indicative of DR or non-DR cases. By distributing the training process across multiple local devices or nodes, individual patient data is kept secure within their respective devices, ensuring privacy is preserved throughout the model training phase. This research not only highlights the prospective artificial intelligence models with deep learning capabilities in the timely identification of DR but also underscores the importance of incorporating privacy-preserving techniques, such as Federated Learning, in healthcare applications. The achieved accuracies of 88.78%, 82.85%, and 83.67% by the Vision Transformer Model, Federated Averaging Model, and Federated Proximal Model, respectively, demonstrate promising results in the development of robust systems for the timely identification of diabetic retinopathy while upholding patient privacy.