Search

Diabetic retinopathy is a leading cause of preventable blindness, with fundus photography commonly used for early detection and severity grading, while deep learning models have shown strong performance in classification but require large, diverse multi-center datasets that are difficult to obtain due to privacy and regulatory restrictions. Because fundus images are protected health information, hospitals cannot share data, resulting in isolated datasets that limit model generalizability across different populations, imaging devices, and clinical settings. To overcome this limitation, a hybrid framework combining federated learning with homomorphic encryption is proposed, allowing multiple hospitals to collaboratively train a shared model without exchanging raw images or plaintext gradients. Each institution performs local training and transmits only encrypted model updates to a central server for secure aggregation, ensuring that patient data remains fully protected while still enabling global model improvement. This approach also mitigates gradient leakage and reconstruction attacks, supports compliance with regulations such as HIPAA and GDPR, and enables scalable, fault-tolerant deployment across heterogeneous healthcare systems, ultimately providing a privacy-preserving pathway for robust multi-center diabetic retinopathy detection.