Develop a method for consistent inter-subject subdivision of segmented streamline tractography results allowing more robust comparisons of tissue microstructure.

Imaging/Tractography: 42 direction, 4 b0, b = 1300s/mm², 2.4mm isotropic resolution DW-data. Subjects (n=15, right handed) were imaged once while depressed and once after remission. Diffusion tensor tractography^2,3 was performed with 45^o/0.2 angular/FA thresholds, 1mm step size and 2mm isotropic seeding.

Subdivision: Pre-segmented tract bundles are first spatially normalised (affine transformation⁴ to an MNI tempate) and, where necessary, individual streamline knot-point orders are reversed to create a uniform direction of streamline propagation. We then calculate (i) an average streamline within each subject and (ii) an average streamline across all subjects (the 'skeleton'), all of which are re-parameterised to N knot points (resulting in N tract subdivisions). To process an individual subject we co-register⁵ the skeleton to their average streamline and, for each streamline in the segmented bundle, find a point, p, with the minimum euclidean distance to any point, a, on the co-registered skeleton. Proceeding 'up' the streamline, such that P = p, p+1, p+2 (and so on), we compare P to A (where A is initially a) and A+1, assigning P to the closest point. If that point was A+1, we increment A before evaluating the next P. Once the end of the streamline is reached, the process is repeated (from p and a) in the opposite direction. After all streamlines are processed we then calculate, for each knot-point on the skeleton, the mean parameter across the set of assigned P.

Incomplete Reconstructions: The above procedure is only effective for subjects with complete tract reconstructions. Where that is not the case, it is therefore necessary to estimate which portion of the tract is present and compensate for the missing sections prior to the subdivision process. Using complete reconstructions we first repeat the above procedure to generate subject wise average streamlines. For each knot point on the set of average streamlines, we then train a fuzzy classifier, providing, within the standardised space, the probability of a voxel being intersected by a correctly formed average streamline at that knot point. Returning to the incomplete reconstruction, those streamlines are spatially normalised as before and a subject wise average streamline calculated. We then interrogate the subjects average streamline against the standard space classification result to determine the most likely set of knot points it represents with respect to a properly formed reconstruction (almost always equivalent to the most likely classification of the local averages start and end points). Finally, the skeleton streamline is then trimmed to the indicated knot-point range, co-registered to the subject average and the point-wise assignment procedure is implemented as above.

Figure 1 displays sub-divisions of the genu of the corpus callosum. Coloured bands correspond to groups of data points assigned to the same knot point along the global average streamline. Figure 2 displays results of the procedure to handle incomplete tracts. Note the locations of coloured bands matches those in Figure 1. Figure 3 provides examples of corticospinal tract, through which (Figure 4) we found significant localised post remission reductions in mean diffusivity.We have demonstrated a method for consistent inter-subject sub-division of streamline bundles and, by doing so, providing localised tract specific parameter estimates. The procedure builds upon the strength of TBSS in three ways. Firstly, by averaging multiple points to provide an estimate, the proposed method is likely to be more robust to noise/image corruption. Secondly, there is no guarantee that the highest/lowest values found by the radial search will be the most important, therefore, perhaps consideration of the wider parameter distribution (as is the case with this method) might provide more representative results. Finally, the radial search is not explicitly constrained to the tract of interest and, as such, where multiple tracts are in close proximity could it potentially sample data points from the incorrect structure – losing tract specificity. By sampling along specifically segmented streamlines (an easily automated task^6,7) tract specificity is guaranteed, thus providing an additional layer of robustness. Turning to the result in Figure 4, there are well-known correlations between depressive state and physical activity⁸, a reduced MD is therefore likely indicative of a 'strengthening' of the motor pathways as subject activity increases post remission; why this would be constrained to the superior portion is currently unknown, but localised microstructural changes in the corticospinal tract in other disease states have been reported previously⁹.