iiinova — where your digital self keeps evolving

Huge thanks to the incredible team of 50 co-authors who made this possible! Special thanks to our lab replicator + dMRI guru @cieslakmatt.bsky.social, the UMN team led by @drdamienfair.bsky.social, and my incredibly supportive postdoc advisor @ted-satterthwaite.bsky.social! Preprint: bit.ly/3QMLG0K

biorxiv.org

032

Takeaway #4: These data are openly available! To lower barriers, we release ABCC on NBDC/LASSO: - >24,000 processed dMRI scans - advanced microstructural metrics - tractography + tabular summaries Together, this is a major new resource for reproducible developmental neuroimaging.

Takeaway #3: Image quality is more complicated than it seems. Some commonly used QC covariates can bias developmental inference. If you are using advanced measures of microstructure (e.g., ICVF) it may not be necessary to include an image quality covariate in analyses.

Takeaway #2: Not all dMRI measures of microstructure are the same! Measures from advanced multi-shell models are more sensitive to development, more consistent with each other, and more robust to noise than tensor-derived measures. These should be prioritized in analyses!

Takeaway #1: Harmonization is essential in multisite dMRI data! Scanner effects are large and spatially structured; they can distort developmental patterns and reduce generalizability unless addressed carefully.

In ABCD, acquisition batch, age, and quality are partially aligned with one another. Given this collinearity, we found that including these age- and batch-aligned quality covariates can attenuate true developmental effects without mitigating noise in the data.

We compared automated QC metrics vs extensive manual ratings (😮‍💨) To our surprise (and joy!), dMRI contrast explains more microstructural variance than expert ratings. Automated QC can outperform visual inspection at scale – saving hours of manual inspection.

How sensitive are dMRI metrics to quality? FA was most susceptible to image quality. Notably, FA was most related to dMRI contrast, and NOT motion. In contrast, advanced metrics like ICVF were overall much more robust to image quality – another plus of moving beyond the tensor.

Different scanners tell the same story; but what about different diffusion metrics? Advanced metrics like ICVF, RTOP, and MKT, exhibited high convergence (ρ ≥ 0.93) in their spatial patterns of development. However, tensor metrics like FA and MD were less consistent.

Do these developmental effects replicate across scanners? Before harmonization, only modestly. After harmonization: near-perfect correspondence across vendors.

141

With our harmonized data, we asked: which metrics best capture development? Advanced dMRI metrics (ICVF, MKT, and RTOP) showed much stronger age effects (>3x!!!) than traditional tensor metrics (FA and MD). Bottom line?: Metric choice strongly impacts sensitivity to development.

141

Given these scanner differences, harmonization is essential! In unharmonized data, acquisition batch explained up to 70% of microstructural variance. We used cutting-edge longitudinal nonlinear harmonization, which eliminated these effects while preserving developmental effects.

162

ABCD has > 60 scanner batches (unique devices / software versions)!!! There were large differences in quality between scanner vendors (GE, Siemens, and Philips). In GE, image quality was related to software version, which itself was related to age! More on that later…

152

Data are distributed in the ABCD-BIDS Community Collection (ABCC). Derivatives include preprocessed images, over 30 microstructural maps from 4 software tools, 60+ white matter bundles, tidy tabular summaries of bundle-wise measures, and 40 automated image quality metrics.