SiMix: A domain generalization method for cross-site brain MRI harmonization via site mixing

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Brain magnetic resonance imaging (MRI) is crucial for disease diagnosis but suffers from cross-site variations due to differing hardware and parameters.
• Existing MRI harmonization methods require training data from all sites, failing when new, unseen sites are introduced.

Purpose of the Study:

• To develop a domain generalization method for brain MRI harmonization that addresses data from unseen sites.
• To improve the consistency and reliability of downstream image processing and analysis across diverse MRI acquisition sites.

Main Methods:

• Proposed Site Mix (SiMix), a novel domain generalization technique for brain MRI harmonization.
• Implemented stochastic mixing of training images from different sites to increase data diversity.
• Introduced a multiview strategy at test time, perturbing unseen site images and ensembling results for enhanced harmonization.

Main Results:

• SiMix demonstrated improved brain MRI harmonization performance for unseen sites.
• The method also showed benefits for harmonizing data from existing sites.
• Experiments on SRPBS and MUSHAC datasets validated the effectiveness of SiMix.

Conclusions:

• SiMix effectively addresses the challenge of brain MRI harmonization for unseen sites using domain generalization.
• The proposed approach enhances the robustness and applicability of MRI harmonization in real-world clinical scenarios.
• SiMix offers a promising solution for consistent cross-site MRI analysis.