KbFL-XAI: Explainable Knowledge-Based Federated Learning for Eye Disease Diagnosis

This paper presents a federated learning approach for detecting diabetic retinopathy using data from four distributed institutions. The method builds a robust model while addressing critical issues of data security, access rights, and privacy protection. Based on Vision Transformer architecture for classifying Diabetic Retinopathy (DR) and normal cases, the model overcomes traditional centralized AI training challenges — efficiency and privacy concerns from data silos — by enabling machine learning without transferring data to a centralized entity. The study investigates the robustness of different FL strategies on diverse data distributions and data quality for ocular diagnosis using retinal fundus images.