Evaluating Transferability of ComBat Harmonization of Diffusion Tensor Magnetic Resonance Imaging Data

Traditional use of the ComBat data harmonization method (a popular means of harmonizing multisite MRI data) is limited by the requirement that the harmonization must be recomputed with any addition of new subjects to the data pool. The goal of this study was to assess whether a transferable ComBat (T-ComBat) algorithm could be applied such that harmonization parameters computed using ComBat on a fixed set of diffusion tensor (DT) MRI training subject data could be reapplied to harmonize new, previously unseen subject data without alteration of the harmonized training data. Emphasis was given to evaluating the necessary size of such a training data pool.

Fractional anisotropy (FA) and mean diffusivity (MD) maps for 314 adolescents were harmonized across two MRI scanning sites using the T-ComBat method applied to new subject data for a variety of training group sizes. For each training group size, FA and MD maps were assessed for differences across sites after T-ComBat on voxel-wise and region-of-interest-wise (ROI) levels. Voxels and ROIs were tested for significant differences across sites.

T-ComBat yielded improved harmonization across sites but did not reach the performance of full ComBat. A limited number of new subjects (approximately 25% for FA, 10% for MD) could be harmonized via T-ComBat and still yield a sufficiently harmonized total dataset.

T-ComBat may be applied to harmonize DT-MRI data from a limited number of new subjects from previously seen scanners without necessitating the reharmonization of previously analyzed data.