Introduction

The Adolescent Brain Cognitive Development (ABCD) study^1,2 is the largest long-term study of brain development and child health in the United States. This consortium has collected neuroimaging, genetic, and behavioral data from 22 sites across the country³ for ~12,000 children ages 108-131 months. Although the study ensured similar imaging protocols across sites, it is not advisable to naively pool these datasets from all sites for joint analysis without additional processing⁴. This is because scanner-related differences can induce significant bias, especially in diffusion MRI (dMRI) data⁵. This bias can be due to different hardware specifications, different acquisition settings, and/or different software versions. Thus, the dMRI data can vary not only between vendors but even between identical scanners running different software versions⁶. The slightest scanner differences therefore can add large unwanted scanner-related effects to the dMRI data, thereby dramatically reducing statistical power. Taking steps to remove these scanner-related effects is thus critically important for neuroimaging studies of psychiatric disorders, where the effect sizes of the group differences related to biology or disease are usually very small⁷. This abstract presents our preprocessing and harmonization strategies on multi-site dMRI data of the ABCD study as part of the NIH R01MH119222 (PIs: Rathi, O’Donnell).

Methods

Dataset.
We worked on the baseline scans of the minimally preprocessed dMRI data from the ABCD study (release 3). Diffusion MRI data was acquired using 3T Siemens, Philips, and GE scanners (Figure 1) from 22 sites using 45 different scanner-related settings in total (Figure 2). Minimal preprocessing steps included: eddy-current correction, motion correction, B0 inhomogeneity correction, gradient unwarp, replacement of bad slice-frames, between scan motion correction, rigid body registration to the atlas, and resampling to 1.7mm^3. See Hatton et al.⁸ for more details of the preprocessing.

Brain Masking.
Data made available does not provide brain masks, which is a prerequisite for our harmonization algorithm. Therefore, we developed a deep learning-based brain masking tool that robustly extracts the brain from dMRI data in a few seconds on a GPU (https://github.com/pnlbwh/CNN-Diffusion-MRIBrain-Segmentation). The network was trained on the dMRI data of 1500 subjects which were collected from multiple clinical and non-clinical datasets. It was tested and manually quality controlled over a thousand unseen subjects from other studies as well as on the ABCD data. We deployed the brain masking in the ABCD dMRI data of 11,622 subjects.

Harmonization.
We next applied our retrospective harmonization algorithm⁴. Our harmonization approach (https://github.com/pnlbwh/dMRIharmonization), based on rotation invariant spherical harmonics (RISH), is applied at the signal level after the minimal preprocessing to remove the scanner differences across multi-site dMRI data, while accounting for non-linearities both in the dMRI signal itself and the scanner bias⁴. After harmonization, any model fitting or connectivity analysis can be applied, unlike other approaches that only harmonize diffusion measures⁹. The performance of this algorithm was also validated on multiple datasets, including a large schizophrenia database¹⁰ as well as in the harmonization community challenge¹¹. Since each site had scanner upgrades or used multiple scanners during the study, we, therefore, applied the harmonization across all 45 scanner-related settings, each of which we term as “dataset” henceforward (Figure 3). In the first step of the harmonization, we selected one dataset as the reference from 45 datasets. The rest of the 44 datasets are referred to as target datasets. A subset of 30 healthy subjects, who were age- and sex-matched at the group level across each target and reference dataset, were selected to learn the scanner differences using RISH features. After that, the learned mappings across RISH features were applied to the entire dMRI data of the corresponding target dataset.

Results

Brain Masking.
The average performance of the algorithm over the unseen thousand subjects, which was not part of the training (of note, ABCD study was not part of this analysis), is over 97% for the Dice overlap score. In the ABCD study, the performance increased to ~99% (compared to manually curated masks). This is most probably due to the fact that 100 dMRI data from the ABCD study was part of our training set (Figure 3).

Harmonization.
As noted, we selected one dataset as a reference, and all other datasets (from the rest of the 44 datasets) were selected as target datasets to be harmonized to the reference independently. Refer to Figure 4, where we demonstrated the performance of the harmonization for two datasets (dataset 1 and dataset 2). To evaluate the performance of the harmonization, average fractional anisotropy along the whole brain white matter was computed for the reference, as well as target datasets before and after harmonization. Unpaired t-tests showed large differences between reference and target datasets (p=0.0017;t=3.52) prior to harmonization. Statistical differences across datasets were removed after harmonization (p=0.53;t=1.15).

Discussion and Conclusion

We presented our brain masking and harmonization efforts on the dMRI data of the ABCD study, collected from 22 sites with 45 datasets, where the harmonized data will be shared in the NIMH data archive. This will allow large-scale data analysis as if the data came from the same scanner, which should significantly increase statistical power with the ability to better characterize the neurodevelopmental changes in the white matter of adolescents.

Acknowledgements

The authors have been supported by the following grant: NIH R01 MH119222 (Rathi, O'Donnell). The project also acknowledges that the research has partly been supported by the BWH Program for Interdisciplinary Neuroscience, through a gift from Lawrence and Tiina Rand (Cetin-Karayumak).