Privacy-Preserving and Collaborative Federated Learning Model for the Detection of Ocular Diseases

Ocular diseases significantly impact the health of the public globally. According to the World Health Organization (WHO) reports, at least 1 billion people suffer from near or distance vision impairment that could have been prevented or has yet to be addressed. These conditions cause difficulty in living a healthy lifestyle and impair individual quality of life. The article explores the application of federated learning in detecting two vision-threatening ocular diseases- diabetic retinopathy and diabetic macular edema. A federated learning framework enhances the technological capabilities of artificial intelligence that leverages decentralised data sources without creating data banks to maintain privacy. The methodology implements a federated learning environment with 2, 3, and 4 clients, using MobileNetV2 as the backbone deep learning model. The model is trained on a composite of 2 datasets procured from the Kaggle repository, comprising coloured fundus images labelled for diabetic retinopathy, diabetic macular edema, and normal cases. The federated learning process involves training at the client end to build client models called local models. The clients in a federated learning system only share updates regarding their local models. The original data is never shared with a central server. The server integrates these local models into the central global models using aggregation strategies such as FedAvg, FedProx, etc. Performance metrics, including prediction accuracy, class-wise accuracy, precision, recall, and F1 score, are calculated across 30 communication rounds. The results demonstrate that the federated learning model achieves an average prediction accuracy of 96%, and a class-wise accuracy of 100% in detecting diabetic macular edema and diabetic retinopathy. The high performance of the federated learning system highlights the significance of federated learning as a viable solution for ocular disease detection while ensuring data privacy.