Introduction

The Adolescent Brain Cognitive Development (ABCD) study is collecting a large repository of data pertaining to adolescent brain development containing cognitive, behavioral, and neuroimaging data from 11,000+ children acquired across 21 sites. Diffusion MRI (dMRI) data acquired as part of the ABCD study is a rich resource to investigate subtle neurodevelopmental pathologies in the tissue microstructure¹. However, dMRI data is particularly sensitive to multi-site/multi-scanner differences due to magnetic field inhomogeneities and vendor-specific data reconstruction algorithms, which introduce substantial bias in the acquired data². Hence, the data must be harmonized prior to any analysis. This abstract presents our harmonization efforts on this very large dataset, building on our success in harmonizing a large schizophrenia database³ as well as in the harmonization community challenge⁴. We validated our harmonization approach using several dMRI measures including Fractional Anisotropy (FA), Return-To-Origin Probability (RTOP), Return To the Axis Probability (RTAP), Return To the Plane Probability (RTPP), and Mean Squared Displacement (MSD)⁵.

Methods

Dataset and brain masking.
We used minimally preprocessed baseline dMRI data from the ABCD study (release 3) acquired on 3T Siemens, Philips, and GE scanners (Figure 1) from 21 sites using 45 different scanner-related settings (Hatton et al.⁶). The ABCD study does not provide the brain masks for the dMRI data, which is a prerequisite for our harmonization. Therefore, we applied and validated our deep learning-based brain masking tool that robustly extracts the brain from dMRI data (https://github.com/pnlbwh/CNN-Diffusion-MRIBrain-Segmentation).

Harmonization.
We applied our retrospective harmonization algorithm² to remove the scanner-related differences across all sites. Our harmonization approach (https://github.com/pnlbwh/dMRIharmonization), based on rotation invariant spherical harmonics (RISH), works at the signal level to remove the scanner differences across sites while accounting for non-linearities both in the dMRI data and the scanner bias. Since each site had scanner upgrades or used multiple scanners during the study, we applied the harmonization across all 45 scanner-related settings, each of which we term as “dataset” henceforward. In the first step of the harmonization, we selected one dataset as the reference and the remaining 44 datasets were harmonized to this reference. To this end, a subset of 35 healthy subjects, who were age-, sex- and IQ-matched at the group level to the reference dataset, were selected to create the RISH feature templates to learn the scanner differences. This step is called template creation. After that, the learned mappings across RISH features were applied to the remaining dMRI data of the corresponding dataset to be harmonized to the reference (see Cetin Karayumak et al.² for details).

Analysis.
We estimated FA, RTOP, RTAP, RTPP, and MSD using DIPY software⁵ on the original and harmonized datasets as well as the reference dataset. We computed the average over the whole brain white matter skeleton and 42 white matter bundles (see Figure 5)^7,8 of these diffusion maps for each subject and compared the original and harmonized datasets to the reference dataset using unpaired t-tests. We set up two experiments to evaluate the performance of the harmonization: (Experiment i) We compared the original, harmonized, and reference datasets that were used in template creation (see Figures 2, 3, and 5a); (Experiment ii) Using an independent set of 35 subjects that were not used in the template creation, we repeated experiment i to demonstrate the performance of the harmonization on unseen data (Figure 4 and 5b).

Results

To-date we have harmonized data from ~10,000 subjects (GE, Siemens scanners). A dataset scanned with Siemens Prisma scanner from Site 16 was selected as a reference due to its large sample size and data quality. Other dMRI datasets with different scanner settings were harmonized to the reference individually.
(Experiment i). Figure 2 compares the average FA across original (before harmonization), harmonized, and reference datasets for six datasets. The FA was computed over the whole brain white matter skeleton and computed independently for each b value (500, 1000, 2000, 3000 s/mm^2). Figure 3 shows the comparison of the average RTOP, RTAP, RTPP, and MSD measures over the whole brain white matter skeleton for dataset 14 (from GE scanner) and dataset 10 (from Siemens Prisma fir scanner). To better understand the regional pattern of differences, we also compared the diffusion measures in 42 white matter bundles across three groups (harmonized, before harmonization, and reference) (e.g., see Figure 5a for the results of RTOP).
(Experiment ii). Figure 4 and Figure 5b show the comparisons using new and unseen subjects of dMRI datasets, which were again age-, sex- and IQ-matched but were not used in the template creation step. Both experiments demonstrated that harmonization removes the existing global and local scanner-related differences across datasets in both single-shell (FA) and multi-shell measures (RTOP, RTPP, RTAP, MSD). Further, it also showed robust performance on unseen new datasets which were not part of the learning.

Discussion and Conclusion

We presented our large-scale harmonization efforts on the multi-site dMRI data from the ABCD study. We note that the harmonized data will be shared in the NDA soon. This will allow large-scale data analysis as if the data comes from the same scanner, which significantly will increase statistical power with the ability to characterize better the neurodevelopmental changes in the white matter of adolescents.

Acknowledgements

The authors have been supported by the following grants: NIH R01 MH119222 (Rathi, O'Donnell) and the BBRF NARSAD Young Investigator Award (PI: Cetin-Karayumak).