2.4.1. Effectiveness of RISH harmonization

Here we assess if acquisition-related differences in diffusion metrics between sites can be removed by RISH harmonization. The first step was to select Training Controls from every cohort that were as similar as possible between sites to minimize sources of variability other than scanner. The Utrecht1 cohort was used as a reference site because the age range of the controls allowed matching with all other sites. This was done on a site-by-site basis, generating four sets of Training Controls with participants from every site matched for age and sex to participants from Utrecht1 (demographics in supplementary Table S1). Tract-based-spatial-statistics (TBSS, Smith et al., 2006) and voxel-based analysis were used to compare the FA and MD between Training Controls of the reference and target sites, before and after harmonization. For the TBSS pipeline, FA and MD were estimated using the diffusion tensor model (dtfit from FSL). Next, FA maps were aligned to the MNI152 template and a white matter skeleton representing the centers of major bundles was computed. Subsequently, FA and MD of the skeleton were compared between reference and target sites in a voxel-wise fashion using t-tests with threshold-free cluster enhancement (5000 permutations). This comparison was also extended to the whole brain to ensure that acquisition-related differences in grey matter regions and other structures are also removed.

We also evaluated the generalizability of effectiveness of harmonization beyond the Training Controls. This was done by creating a group of Validation Controls with data from Utrecht1 (n = 15), Munich (n = 15) and Singapore (n = 15), since those sites had a sufficient number of controls outside the Training Controls to generate separate sets of matched groups (demographics in supplementary Table S2). Similar to the analysis with training controls, TBSS and voxel-based analysis were used to compare validation controls between each target site and the reference, before and after harmonization. Furthermore, one-way ANOVA was performed to compare the average FA and MD of the TBSS skeleton across these three sites.