Relay Learning: A Physically Secure Framework For Clinical Multi-site Deep Learning

This research evaluated Relay Learning, a flexible and secure multi-site deep learning framework, in different multi-site clinical problems for different diseases and anatomical structures. The framework enables clinical deep learning models to fuse the broad knowledge offline from multi-site data without privacy leakage, and was tested to segment important retina fundus structures on several widely used public datasets, and to develop a multi-site simulation system for mediastinum tumor diagnosis. Additionally, it was deployed in five medical institutions and trained a deep model to perform brain midline shift detection.