Federated learning for medical image analysis: A survey
Hao Guan1, Pew-Thian Yap1, Andrea Bozoki2
1Department of Radiology and Biomedical Research Imaging Center, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, USA.
Pattern Recognition
|April 1, 2024
Summary
Federated learning addresses small sample sizes in medical imaging by enabling collaborative model training across sites without sharing sensitive patient data. This survey reviews methods, challenges, and opportunities in this privacy-preserving approach.
Area of Science:
- Medical Imaging
- Machine Learning
- Artificial Intelligence
Background:
- Machine learning in medical imaging is hindered by small sample sizes.
- Sharing data across institutions for larger datasets is restricted by privacy concerns.
- Federated learning offers a solution for collaborative model training without data sharing.
Purpose of the Study:
- To provide a comprehensive survey of federated learning methods in medical image analysis.
- To systematically review research published between 2017 and 2023.
- To identify current challenges and future research opportunities.
Main Methods:
- Systematic literature review across major scientific databases (IEEE Xplore, ACM, PubMed, etc.).
- Categorization of federated learning methods based on client-end, server-end, and communication techniques.
- Review of benchmark datasets, software platforms, and an experimental evaluation of typical methods.
Main Results:
- Federated learning methods are categorized and analyzed based on system architecture and specific medical imaging problems.
- Existing datasets and software tools for federated learning in medical imaging are reviewed.
- An empirical evaluation of common federated learning techniques is presented.
Conclusions:
- Federated learning is a promising solution for privacy-preserving collaborative machine learning in medical imaging.
- The survey highlights current research status, challenges, and potential avenues for future work.
- Understanding different federated learning approaches is crucial for advancing medical image analysis.
