Federated learning for medical imaging radiology

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Healthcare
• Medical Imaging Analysis

Background:

• Federated learning (FL) is increasingly adopted in medical AI for its potential in enhancing model accuracy, ensuring patient privacy, and enabling generalisability across institutions.
• Despite its promise, FL research specifically for medical imaging AI is nascent, with a notable divergence between theoretical advancements and practical clinical implementation.

Purpose of the Study:

• To review current research in federated learning for medical imaging AI.
• To delineate the distinctions between state-of-the-art (SOTA) research and state-of-the-practice (SOTP) applications.
• To identify future research trajectories for translating SOTA findings into SOTP clinical realities.

Main Methods:

• Systematic review of recent literature on federated learning in medical imaging.
• Comparative analysis of published research (SOTA) versus applied clinical research (SOTP).
• Identification of key factors influencing the translation from SOTA to SOTP.

Main Results:

• Significant gap identified between published federated learning research and its application in clinical settings.
• Key challenges in data handling, model development, system architecture, governance, and human-in-loop integration were highlighted.
• Future research directions are proposed to facilitate the transition of FL from theoretical to practical medical imaging applications.

Conclusions:

• Federated learning holds substantial promise for medical imaging AI, but practical implementation requires addressing the identified research gaps.
• Translating state-of-the-art findings into state-of-the-practice necessitates focused efforts on data, models, systems, governance, and human-AI collaboration.
• Further research is crucial to realize the full potential of FL for multi-institutional medical AI collaboration and clinical deployment.