Related Experiment Video
Updated: Jan 17, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
MSAFed: Generalized Multi-Stage and Adaptive Federated Learning for Test-Time Medical Segmentation
IEEE Journal of Biomedical and Health Informatics
|September 16, 2025
Summary
Federated learning (FL) models struggle with generalization. MSAFed, a new framework, improves FL model performance both within and beyond participating centers using adaptive strategies and contrastive learning.
Area of Science:
- Artificial Intelligence in Healthcare
- Machine Learning
- Medical Informatics
Background:
- Federated learning (FL) facilitates collaborative AI model training across institutions without data sharing, enhancing privacy in healthcare.
- Existing FL models exhibit poor generalization within federated networks and underperform on unseen data, especially in diverse medical domains.
- Current test-time adaptation techniques for FL do not adequately mitigate biases towards source data distributions, limiting clinical utility.
Purpose of the Study:
- To introduce MSAFed, a generalized multi-stage adaptive federated learning framework.
- To enhance both the internal generalization (across clients) and external test-time adaptation (to unseen clients) of FL models.
- To address the limitations of current FL methods in heterogeneous medical environments.
Main Methods:
- Developed a pretraining strategy incorporating intra-client and inter-client contrastive learning with prototype-aware aggregation to create a generalized global model.
- Implemented an adaptive learning rate strategy to boost generalization within the FL network.
- Utilized source knowledge, including adaptive learning rates and prototypes, for dynamic network adaptation during test time for unseen clients.
Main Results:
- MSAFed demonstrated superior performance on both internal and external FL tasks across three real-world multi-center medical datasets.
- The framework effectively improved generalization across participating clients.
- Enhanced adaptability to unseen clients and heterogeneous domains was achieved.
Conclusions:
- MSAFed offers a robust solution for improving federated learning model performance in healthcare.
- The proposed adaptive framework addresses key generalization and adaptation challenges in medical AI.
- The approach holds significant potential for advancing privacy-preserving AI applications in clinical settings.
