Federated learning for medical imaging radiology

Abstract

Federated learning (FL) is gaining wide acceptance across the medical AI domains. FL promises to provide a fairly acceptable clinical-grade accuracy, privacy, and generalisability of machine learning models across multiple institutions. However, the research on FL for medical imaging AI is still in its early stages. This paper presents a review of recent research to outline the difference between state-of-the-art [SOTA] (published literature) and state-of-the-practice [SOTP] (applied research in realistic clinical environments). Furthermore, the review outlines the future research directions considering various factors such as data, learning models, system design, governance, and human-in-loop to translate the SOTA into SOTP and effectively collaborate across multiple institutions.

Figure 1.

A model’s lifecycle in federated learning environments in realistic clinical settings. (1) the learning process is initiated by an orchestration server, (2) the server selects, registers, and/or clusters participating institutions, (3) the server broadcasts the training and evaluation configurations along with information about datasets, (4) participating institutions execute their local single/multitask trainers and heterogeneous data handlers, (5) participating institutions optionally perform local model enhancements and transmit local models to the server, (6) the server performs aggregation and/optimisation, and/or enhances global model, (7) the server evaluates the global model, (8) the server sends the model for retraining if it does not meet the desired convergence/performance criteria. After the model converges, (9) the server then sends the converged model for versioning and possible deployment at participating institutions, and finally (10) the participating institutions keep monitoring the model performance and either initiate new training cycles or abandon the model if it is not required anymore.