Federated learning for lesion segmentation in multiple sclerosis: a real-world multi-center feasibility study.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Multiple sclerosis (MS) is a chronic, disabling neuroinflammatory disease.
• Accurate segmentation of MS lesions on MRI is vital for disease monitoring and treatment evaluation.
• Manual segmentation is labor-intensive and variable; automated methods require large, diverse datasets, posing privacy and data-sharing challenges.

Purpose of the Study:

• To apply and evaluate Federated Learning (FL) for automated multiple sclerosis (MS) lesion segmentation in a real-world clinical setting.
• To assess the feasibility of using FL to train AI models on distributed datasets without compromising patient privacy.

Main Methods:

• Utilized the self-configuring nnU-Net model within a Federated Learning framework.
• Trained the model on 512 MRI cases from three different clinical sites without direct data sharing.
• Assessed model performance using Dice scores on held-out test sets.

Main Results:

• The federated model achieved Dice scores between 0.66 and 0.80 across test sets.
• Performance varied across sites, indicating the impact of data heterogeneity.
• Demonstrated the potential of FL for scalable and secure AI in distributed neuroimaging.

Conclusions:

• Federated Learning offers a viable, privacy-preserving method for developing robust AI tools for MS lesion segmentation.
• This approach facilitates collaborative AI development in neuroimaging, addressing data privacy and regulatory concerns.
• Supports the adoption of secure, collaborative AI for medical research and clinical applications.