Federated Learning Enhances Neuroimaging Predictions in Multiple Sclerosis

Overview

Federated learning (FL) using the FLightcase toolbox outperformed centralized training in predicting brain age from MRI data and showed promise in predicting cognitive performance in multiple sclerosis (MS) patients. Real-world deployment across three international centers demonstrated feasibility without sharing sensitive data.

Background

Deep learning has revolutionized brain imaging analysis, enabling accurate quantification of brain structures and prediction of biological brain age. However, large datasets are required for reliable models, and data sharing is often restricted by privacy, legal, and ethical concerns, especially with patient data. Federated learning addresses these challenges by training models locally and sharing model parameters instead of data, but real-world applications in neuroimaging remain limited due to technical and logistical barriers.

Data Highlights

Key Findings

• Federated learning outperformed centralized training in predicting brain age from MRI, reducing mean absolute error from 7.02 to 6.08 years.
• Strong correlations between true and predicted age were observed across all centers (ranging 0.88 to 0.93, all p < .001).
• Deep transfer learning for predicting cognitive performance (SDMT) in MS patients was more accurate (MAE 9.19) than shallow transfer learning (MAE 11.05).
• Federated deep transfer learning predicted SDMT with varying accuracy across centers, with MAEs between 8.98 and 10.71 and significant correlations in Greifswald and Prague.
• FL enabled access to large, geographically distributed MS imaging datasets without sharing sensitive patient data.
• Challenges remain in addressing non-independent and identically distributed (non-IID) data and integrating other imaging modalities to improve model performance.

Clinical Implications

Federated learning offers a practical approach to leverage large neuroimaging datasets across institutions while maintaining patient privacy, facilitating improved predictive modeling in MS. Adoption of FL can enhance cognitive performance prediction from MRI without the need for data centralization, potentially accelerating research and clinical decision-making. Future improvements in FL algorithms to handle heterogeneous data will further optimize clinical utility.

Conclusion

This study demonstrates the feasibility and advantages of real-world federated learning in neuroimaging research for MS, providing a scalable framework that preserves data privacy and enhances predictive accuracy. The open-source FLightcase toolbox may catalyze broader adoption of FL in clinical brain imaging studies.

References

1. McMahan et al. 2017 -- Communication-Efficient Learning of Deep Networks from Decentralized Data
2. Lecun et al. 2015 -- Deep Learning
3. UK Biobank 2020 -- Brain Imaging Data
4. FLightcase Toolbox 2023 -- Federated Learning for Neuroimaging