Cross-site harmonization of diffusion MRI data without matched training subjects.

Area of Science:

• Neuroimaging
• Medical Physics

Background:

• Diffusion MRI (dMRI) data exhibit significant cross-site variability, hindering pooled analyses.
• Existing harmonization methods like rotational invariant spherical harmonics (RISH) often require matched healthy control groups, which are not always available retrospectively.

Purpose of the Study:

• To develop and validate a new framework for harmonizing dMRI data that does not require matched training groups.
• To enable effective cross-site variability reduction for pooled dMRI analyses.

Main Methods:

• A voxel-based generalized linear model (GLM) framework, termed RISH-GLM, was developed to learn harmonization features.
• The method controls for potential covariates and harmonizes data from multiple sites simultaneously in a single step.
• Unlike traditional RISH, RISH-GLM does not necessitate matched training subjects.

Main Results:

• RISH-GLM demonstrated equivalence to conventional RISH when trained with matched subjects.
• The framework successfully learned harmonization with highly unmatched subject groups.
• Simultaneous harmonization of data from three different sites was effectively achieved.

Conclusions:

• RISH-GLM offers a flexible approach to dMRI data harmonization, accommodating both matched and unmatched training groups.
• The method enables efficient, single-step harmonization of data from multiple sites, facilitating broader dMRI research.