Privacy-Preserving Federated Learning for Brain MRI Analysis

Advancements in medical imaging facilities have led to a substantial increase in the generation of medical imaging data on a daily basis. This creates an excellent opportunity for researchers to develop data-driven methods to improve healthcare. However, one of the major challenges in this regard is the requirement of large, diverse, and high-quality datasets for effective training, which are often scarce within a single institution. Although combining datasets from multiple institutions into a central repository would be ideal, data privacy regulations pose significant obstacles, particularly when involving institutions from different countries. To address this issue, Federated Learning (FL) can be employed, enabling the training of machine learning models on distributed data from various sites. In this research, federated learning has been employed for Brain MRI Analysis and to assess the generalizability of the models using data from institutions outside the federation. The distributed architecture of FL, however, introduces a vulnerability to adversarial users who may manipulate the global model by modifying their local updates or datasets. To overcome this challenge, FedByz aggregation algorithm has been proposed, which enhances the resilience of FL against Byzantine users. The experiments demonstrate that this approach achieves convergence even in the presence of Byzantine users, and its accuracy is comparable to that of traditional federated learning benchmarks.