To evaluate the feasibility of federated learning (FL) in neuroimaging research, specifically predicting cognitive status in multiple sclerosis (MS) patients using brain MRI data, while addressing data privacy concerns.
Key Findings:
Federated training achieved a mean absolute error (MAE) of 6.08 for age prediction, outperforming centralized training (MAE = 7.02), indicating superior model performance.
Deep transfer learning yielded better performance on SDMT prediction (MAE = 9.19) compared to shallow transfer learning (MAE = 11.05), highlighting the effectiveness of deep learning techniques.
Pearson correlations for SDMT predictions ranged from 0.25 to 0.50 across centers, indicating varying predictive success and suggesting areas for improvement.
Interpretation:
Federated learning is a viable approach for neuroimaging research, allowing access to large datasets without compromising patient data privacy, thus fostering collaboration across institutions.
Limitations:
Challenges remain in addressing non-IID data issues, which can skew model training and performance, and integrating various imaging modalities.
The study's scope was limited to three centers, which may affect generalizability and the applicability of findings to broader populations.
Conclusion:
FLightcase demonstrates the practicality of federated learning in neuroimaging, encouraging further exploration and development of FL methodologies in real-world settings, while emphasizing the need to address existing limitations.
by Stijn Denissen, Jorne Laton, Matthias Grothe, Manuela Vaneckova, Tomáš Uher, Matěj Kudrna, Dana Horáková, Johan Baijot, Iris-Katharina Penner, Michael Kirsch, Jiří Motýl, Maarten De Vos, Oliver Y. Chén, Jeroen Van Schependom, Diana Maria Sima, Guy Nagels
