Real-world federated learning for brain imaging scientists

Area of Science:

• Neuroimaging
• Machine Learning
• Federated Learning

Background:

• Federated learning (FL) offers potential for deep learning in neuroimaging but is underutilized in real-world settings.
• The study introduces FLightcase, a novel FL toolbox specifically designed for brain research.
• FLightcase is evaluated on a real-world FL network for predicting cognitive status in multiple sclerosis (MS) patients using brain MRI.

Purpose of the Study:

• To develop and evaluate FLightcase, a federated learning toolbox for neuroimaging research.
• To assess the feasibility and performance of FL in predicting cognitive status in MS patients using brain MRI.
• To compare federated learning with centralized approaches and evaluate different transfer learning strategies.

Main Methods:

• Trained DenseNet models to predict age from T1-weighted brain MRI across three centers (Brussels, Greifswald, Prague).
• Benchmarked federated model performance against a centralized version.
• Fine-tuned the best brain age model using shallow and deep transfer learning (TL) to predict Symbol Digit Modalities Test (SDMT) performance in MS patients.

Main Results:

• Federated training outperformed centralized training for brain age prediction (MAE 6.08 vs. 7.02).
• High correlations between true and predicted age were achieved with federated learning (r=0.88-0.93).
• Deep TL was superior for SDMT prediction (MAE 9.19) compared to shallow TL (MAE 11.05), with federated deep TL achieving MAEs of 8.98-10.71 across centers.

Conclusions:

• Real-world federated learning with FLightcase is viable for neuroimaging research in MS, enabling large-scale data analysis without data sharing.
• The federated SDMT-prediction model shows promise and can be enhanced by addressing non-IID data challenges and incorporating multimodal imaging.
• The study encourages the adoption of FL in neuroimaging by providing detailed real-world experiments and an open-source toolbox.