Multiple sclerosis lesion segmentation: revisiting weighting mechanisms for federated learning

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• Federated learning (FL) enables collaborative medical image analysis without data sharing.
• FL application in neuroimaging, specifically multiple sclerosis (MS) lesion segmentation, is suboptimal due to scanner and acquisition parameter variations.
• Existing FL methods struggle with the heterogeneity of MS lesion characteristics.

Purpose of the Study:

• To develop the first federated learning framework for multiple sclerosis (MS) lesion segmentation.
• To enhance FL performance in neuroimage analysis by addressing data heterogeneity.
• To improve the accuracy and robustness of MS lesion segmentation using FL.

Main Methods:

• Proposed a novel FL framework incorporating two re-weighting mechanisms for MS lesion segmentation.
• Implemented a learnable weight for each local node during aggregation, based on segmentation performance.
• Introduced re-weighting of the segmentation loss function per client, considering lesion volume during training.

Main Results:

• Validated the framework on public and clinical datasets for FL MS segmentation.
• Achieved case-wise and voxel-wise Dice scores of 65.20% and 74.30% on a public dataset.
• Obtained case-wise and voxel-wise Dice scores of 53.66% and 62.31% on an in-house dataset.

Conclusions:

• The proposed FL method significantly outperforms existing FL approaches for MS lesion segmentation.
• The re-weighting mechanisms effectively address data heterogeneity in federated neuroimage analysis.
• The developed FL framework achieves performance comparable to centralized training, demonstrating its efficacy.