Multi-site harmonization of diffusion MRI data in a registration framework.

Area of Science:

• Neuroimaging
• Medical Physics
• Biomedical Engineering

Background:

• Diffusion MRI (dMRI) data exhibits significant variability across scanners, even with similar acquisition parameters.
• This inter-scanner variability hinders large-scale neuroimaging studies by limiting sample size and statistical power.
• Harmonization of dMRI data is crucial for robust and reproducible research.

Purpose of the Study:

• To present a novel method for harmonizing raw dMRI signal data.
• To address scanner-specific signal differences without relying on derived diffusion measures.
• To enhance the utility of multi-site dMRI datasets for increased statistical power.

Main Methods:

• Developed a harmonization approach using rotation invariant spherical harmonic (RISH) features.
• Integrated RISH features within a multi-modal image registration framework.
• Registered all dMRI datasets to a common template and harmonized signals based on group-level, voxel-wise RISH feature differences.

Main Results:

• Validated the method on dMRI data from seven different scanner sites (GE, Philips, Siemens).
• Demonstrated significant reduction in scanner-specific differences in diffusion measures (FA, MD, GFA) post-harmonization.
• Confirmed preservation of disease-related diffusion abnormalities using synthetic data and independent validation with TBSS.

Conclusions:

• The proposed method effectively removes scanner-specific signal variations in dMRI data.
• This harmonization technique is model-independent and preserves biologically relevant information.
• The approach facilitates increased sample sizes and statistical power in multi-site neuroimaging research.