Synthetic Images for Federated Learning of Medical Images

• Activity Number: 488 - Novel Methods for Unique Spatial Imaging Applications
• Type: Contributed
• Date/Time: Wednesday, August 10, 2022 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Imaging
• Author(s): Tsung-Hung Tsai and Henry Horng-Shing Lu*
• Companies: National Yang Ming Chiao Tung University, Taiwan and National Yang Ming Chiao Tung University, Taiwan
• Keywords: federated learning; machine learning; chest X-ray images; COVID-19; color fundus images; glaucoma

Abstract

Federated Learning (FL) provides the feasible approach to jointly train the machine learning model without accessing private data in various medical institutions. However, the resulting FL model could be biased towards institutions with larger training datasets. Institutions with a relatively limited amount or heterogeneous types of data may have lower accuracy. In this study, we propose the framework of Synthetic Images for Federated Learning (SIFL) that integrates the information of local institutions with heterogeneous types of data. The main concept of the SIFL is to develop the resulting global model that can handle the diversity in heterogeneous types of data collected in local medical institutions by synthetic images similar to minor types in local collections. We use the chest X-ray images of COVID-19 and the color fundus images of glaucoma to compare the results from the centralized learning and several decentralized learning frameworks. The result demonstrates that SIFL-based models outperform the conventional FL models. For institutions with the relatively limited amount of heterogeneous types of data, the model performance of SIFL is improved in these empirical studies.