Harmonizing multisite neonatal diffusion-weighted brain MRI data for developmental neuroscience.

Abstract

Large diffusion-weighted brain MRI (dMRI) studies in neonates are crucial for developmental neuroscience. Our aim was to investigate the utility of ComBat, an empirical Bayes tool for multisite harmonization, in removing site effects from white matter (WM) dMRI measures in healthy infants born at 37 gestational weeks+ 0 days-42 weeks+ 6 days from the Theirworld Edinburgh Birth Cohort (n = 86) and Developing Human Connectome Project (n = 287). Skeletonized fractional anisotropy (FA), mean, axial and radial diffusivity (MD, AD, RD) maps were harmonized. Differences between voxel-wise metrics, skeleton means and histogram widths (5th-95th percentile) were assessed before and after harmonization, as well as variance associated with gestational age at birth and scan. Before harmonization, large cohort differences were observed. Harmonization removed all voxel-wise differences from MD maps and all metric means and histogram widths, however small voxel-wise differences (<1.5 % of voxels) remained in FA, AD and RD. We detected significant relationships between GA at birth and all metrics. When comparing single site and multisite harmonized datasets of equal sample sizes, harmonized data resulted in smaller standardized regression coefficients. ComBat could enable unprecedented sample sizes in developmental neuroscience, offering new horizons for biomarker discovery and validation, understanding typical and atypical brain development, and assessing neuroprotective therapies.