GoKawiilEthics
The research consists of secondary analyses of de-identified primary datasets. Information regarding informed consent of participants (or guardians) in primary studies can be found in the references in Supplementary Table S2.T2. Secondary analysis of these data was also approved by the Vanderbilt University Medical School Institutional Review Board (IRB no. 210968).
Software reporting
All analysis visualizations and statistics were generated using Python v3.9.20. Notable Python libraries used are pandas (v2.2.3), numpy (v2.0.2), seaborn (v0.13.2), matplotlib (v3.9.2), rpy2 (v3.5.11) and scipy (v1.13.1). For creating 3D qualitative visualizations of WM tracts, we used the mrview tool inside the MRtrix3 (version 3.0.4). Alternatively, for qualitative region of interest 3D visualizations, the OpenDIVE visualization library (v0.3.0) was used with a different Python (v3.10.18) for library compatibility purposes. Regarding the image preprocessing and processing software, please refer to ‘MRI processing pipeline’.
Data
We aggregated diffusion-weighted imaging (DWI) and T1-weighted (T1w) data from 50 independent studies spanning 0 to 100 years of age (Extended Data Table 1), encompassing 75,036 DWI scans from cognitively normal and clinical participants. All data were converted from DICOM to NIfTI using dcm2niix and organized in BIDS format in accordance with previous work44. Notably, the 50 included datasets were harmonized through standardized preprocessing, postprocessing, and feature-wise statistical harmonization using GAMLSS.
MRI processing pipeline
DWI data were preprocessed using the PreQual pipeline45, which corrects for susceptibility-induced, motion-related and eddy current distortions, and performs slice-wise signal imputation. Specifically, image denoising was performed using the MRtrix3 toolkit46 (v3.0.4) implementation of Marchenko–Pastur principal component analysis. Following this, TOPUP from the FSL (v6.0.4) software library47 is used for susceptibility-distortion correction. For DWI without reverse phase encoding scans acquired, TOPUP is run using a synthetic b0 image created from a T1w scan from the same imaging session via the Synb0-DISCO48 algorithm (v3.1). FSL’s EDDY is then used for motion and eddy current-distortion correction, also performing slice-wise signal imputation with the flag ‘--repol‘. Following preprocessing, DTI models were fit to all volumes with b-values ≤ 1,500 s mm−2 (refs. 40,49) using dwi2tensor from MRtrix3. DTI-derived scalar maps—including FA, MD, AD and RD—were computed using tensor2metric from MRtrix346.
To enable consistent tract segmentation, all diffusion data were resampled to 1 mm isotropic resolution using MRtrix350. Tractography was performed using TractSeg51 (v2.8), which automatically segments 72 anatomically defined cerebral WM pathways (see Extended Data Table 2 for a list of tracts and abbreviations). For each tract, we computed streamline density-weighted averages of DTI metrics (FA, MD, AD and RD) as well as macrostructural features—volume, length and surface area—using the Scilpy toolkit52 (v1.5.0) scripts scil_compute_bundle_mean_std.py and scil_evaluate_bundles_individual_measures.py, respectively. Both raw and ratio-normalized macrostructural features are provided, where volume, surface area and average length of tracts are normalized to total brain volume (excluding ventricles), estimated total intracranial volume and cerebral WM volume.
T1w images were included only when acquired in the same session as DWI data. Brain segmentation was performed using the recon-all command from FreeSurfer (v7.2.0)53, yielding estimates of cerebral WM volume, brain volume excluding ventricles and total intracranial volume. For participants aged ≤2 years, we employed infant_recon_all from the infant FreeSurfer pipeline54 (v0.0) to account for age-specific brain morphology. Cerebral WM masks were defined using MRtrix346 5TT labels, excluding the cerebellum and brainstem. T1w images and corresponding WM segmentations were rigidly registered to DWI space using FSL’s47 epi_reg. Whole-brain WM DTI metrics were then computed by averaging values within the WM mask.
... continue reading