Abstract
Deriving white matter (WM) bundles in-vivo has thus far mainly been applied in research settings, leveraging high angular resolution, multi-shell diffusion MRI (dMRI) acquisitions that enable advanced reconstruction methods. However, these advanced acquisitions are both time-consuming and costly to acquire. The ability to reconstruct WM bundles in the massive amounts of existing single-shelled, lower angular resolution data from legacy research studies and healthcare systems would offer much broader clinical applications and population-level generalizability. While legacy scans may offer a valuable, large-scale complement to contemporary research datasets, the reliability of white matter bundles derived from these scans remains unclear. Here, we leverage a large research dataset where each 64-direction dMRI scan was acquired as two independent 32-direction runs per subject. To investigate how recently developed bundle segmentation methods generalize to this data, we evaluated the test-retest reliability of the two 32-direction scans, of WM bundle extraction across three orientation distribution function (ODF) reconstruction methods: generalized q-sampling imaging (GQI), constrained spherical deconvolution (CSD), and single-shell three-tissue CSD (SS3T). We found that the majority of WM bundles could be reliably extracted from dMRI scans that were acquired using the 32-direction, single-shell acquisition scheme. The mean dice coefficient of reconstructed WM bundles was consistently higher within-subject than between-subject for all WM bundles and ODF reconstruction methods, illustrating preservation of person-specific anatomy. Further, when using features of the bundles to predict complex reasoning assessed using a computerized cognitive battery, we observed stable prediction accuracies ( r : 0.15-0.36) across the test-retest data. Among the three ODF reconstruction methods, SS3T had a good balance between sensitivity and specificity in external validation, a high intra-class correlation of extracted features, more plausible bundles, and strong predictive performance. More broadly, these results demonstrate that bundle segmentation can achieve robust performance even on lower angular resolution, single-shell dMRI, with particular advantages for ODF methods optimized for single-shell data. This highlights the considerable potential for dMRI collected in healthcare settings and legacy research datasets to accelerate and expand the scope of WM research.