Automatic scan range for dose-reduced multiphase CT imaging of the liver utilizing CNNs and Gaussian models

Ha Manh Luu, Theo van Walsum, Hong Son Mai, Daniel Robert Franklin, Thi Thu Thao Nguyen, Thi My Le, Adriaan Moelker, Van Khang Le, Dang Luu Vu, Ngoc Ha Le, Tran Quoc Long, Trinh Chu Duc, Nguyen Linh Trung

Abstract: Highlights•We propose a novel generative replay technique to address the challenge of data heterogeneity among institutions that participate in collaborative learning. As opposed to existing methods that either provide sophisticated ways to control the optimization strategy or share partial global data to mitigate the performance drops from data heterogeneity, our generative replay technique provides a new insight for the development of collaborative learning methods, which is easy to implement, and can be applied to numerous types of deep learning tasks (e.g., classification, regression, etc.).•Our generative replay technique is flexible to use. It can either be incorporated into existing federated learning framework to increase their capability of handling data heterogeneity across institutions with minimal modifications, or be used as a novel and individual collaborative learning framework to reduce communication cost and mitigate privacy cost.•While previous collaborative learning methods require frequent communication between local institutions and the central server, the proposed FedReplay only requires a one-time communication between local institutions and the central server, which is time-efficient. The training of primary model is performed solely on the central server, without restriction to hardware and network speeds among institutions.