Decentralized Diagnostics: The Role of Federated Learning in Modern Medical Imaging

The digitization of medical diagnostics poses significant challenges regarding data privacy and the efficient management of decentralized data sources. Decentralized diagnostics examines the transformative potential of federated learning in the medical imaging field. Federated learning can train models across multiple data nodes without centralized data aggregation, addressing important data privacy issues while optimizing the utilization of widespread data repositories. This chapter discusses the symbiotic relationship between medical imaging and federated learning, explores current applications, and explains technical mechanisms. Most federated learning research focuses on perspectives like communication efficiency, privacy protection, and personalization. However, data annotation demands expertise and tedious work in the medical image analysis scenario. The federated active learning framework can reduce the annotation workload while maintaining algorithm performance on medical images. Despite using only 50% of the sample, our proposed framework can achieve high accuracy in real-world dermoscopy tasks, outperforming comparable state-of-the-art performance for complete data. In addition, the main challenge of deep learning in computer-assisted diagnosis is to collect large heterogeneous datasets from multiple hospitals to build powerful deep learning models due to hospitals’ strict regulations for sharing sensitive medical data. However, some previous federated learning methods have problems with high computational demands and a lack of data at the local hospital level. To address this problem, we developed a federated learning method combined with a reasonable deep learning model, a massive-training artificial neural network (MTANN), in tumor segmentation in CT. The method was applied to segment combined liver tumors on a small dataset. Our algorithm achieved a dice of 0.712, comparable to 0.723 of the gold standard centralized training model, which was higher than 0.592 (P < 0.05) by the advanced Res-U-Net model. Moreover, the proposed model required 0.25 h of training time, which was less than 5 h of Res-U-Net on a server GPU.