Tractometry combines diffusion MRI metrics along specific white matter bundles extracted via tractography.¹ It expands on voxel-based morphometry (VBM)² performed on a voxel-per-voxel level of the white matter and tract-based spatial statistics (TBSS)³ performed on the voxels of the white matter skeleton. Hence, before using tractometry in applications, we must confirm whether its processing steps and subsequent results are reproducible. In this test-retest study, we validate that fiber bundles are reproducible in terms of their shape, volume and average metrics when these bundles are considered as a single ROI. Then, we split each bundle into 20 equidistant parts and study the reproducibility of diffusion metrics.⁴ Results are crucial to consider for future study design of tractometry applications.

Diffusion-weighted images were acquired on 11 healthy subjects with 3 different acquisitions each along 64 uniformly distributed directions using a b-value of 1000 s/mm² with 2 mm isotropic resolution.

Our processing pipeline is illustrated in Figure 1. Fiber ODFs and DTI/HARDI metrics are first extracted from the raw dMRI data using Dipy (a)⁵ (only FA will be reported in this abstract). Then, whole brain fODF tractography is performed using particle-filter tracking with anatomical priors (b).⁶ Next, 30 white matter bundles (c) are automatically dissected from the whole brain tractogram with TractQuerier.⁷ From this point on, each bundle is processed independently using our tractometry pipeline (d). First, short and long streamlines are pruned-out based on user-defined bundle dependent thresholds. Second, spurious streamlines (outliers) are removed with hierarchical QuickBundles (e).⁸ Third, centroids are computed as a mean streamline of the bundle using the minimum-distance-flipped metric (f).⁹ This centroid is subsampled on N=20 equidistant points and every point of every streamline of the bundle is assigned to the closest centroid point (g). With these assignments, it is possible to extract tract profiles for every metric of interest as well as a single average metric for each bundle of interest (h).

Our first experiment was to validate the shape and volume of our bundles. We first used registration to put all bundles in the same space to make sure they overlapped. Linear and nonlinear diffeomorphic registration was performed on the T1s via ANTs registration to the MNI 2009 template.^10,11 The resulting warps were then applied to the corresponding bundles with nearest-neighbor interpolation. Once these volumes were in the same space, Dice’s coefficient, which measures the overlap of two volumes across intrasubject and intersubject acquisitions, was computed.¹²

Afterwards, we tested the reproducibility of the metrics. By simply looking at the average value of a metric for an entire bundle, we computed a percentage difference for all possible pairs of acquisitions and checked whether intrasubject acquisitions differed from intersubject acquisitions.

Finally, by looking at the tract profiles, we computed the metric values of all 20 parts across all bundles for each acquisition and the variability with respect to the position in the bundle.

Figure 2 illustrates the average Dice’s coefficient for every bundle. We see that bundle volumes are about 10% closer when compared to the same subject than to other subjects, while still having a decent overlap overall.

We also provide a 33x33 similarity matrix (c.f. Fig. 3) in which each entry [i,j] contains the mean FA difference between acquisition i and j. The three intrasubject acquisitions being next to each other, we can clearly see that they are much closer together (in blue) than to others (in red). The maximal difference is only 9% since we are averaging the FA over the whole bundles.

Figure 4 shows the distribution of FA values along the tract profile of the left corticospinal tract (CST) for a single acquisition, and an average over all 33 acquisitions. We can clearly see that the FA values of the single subject are well within the standard deviation of all acquisitions. This observation also holds true for every other bundles.

We have shown that tractometry creates tract profiles of bundles that have similar overlapping volumes, have metric values closer to the same subjects than to others, and have a consistent trend when split into 20 parts. Results specifically on the left CST and the FA metric were featured, but similar satisfying results were obtained with other combinations. In the future, machine learning techniques will be applied to these bundle-metric combinations on multiple populations in order to find significantly discriminant regions of the brain.