Mapping connectomes with diffusion MRI: Deterministic or probabilistic tractography?

In the Discussion section of the manuscript entitled “Mapping connectomes with Diffusion MRI: deterministic or probabilistic tractography?”,1 Sarwar and colleagues briefly mention a class of methods that aim to imbue a tractogram reconstruction with quantitative properties, by modulating the contribution of individual streamlines in such a manner that the tractogram becomes concordant with the underlying image data.2-5 They state that, while such methods “can potentially improve the accuracy of connectome reconstruction”, one particular method (Spherical‐deconvolution Informed Filtering of Tractograms, SIFT)2 “can potentially yield spurious between‐group differences” when the densities of specific pathways are compared quantitatively across subjects. They additionally present a hypothetical fiber phantom intended to support this statement. Here we demonstrate that the empirical behavior suggested to be problematic by Sarwar and colleagues is in fact the best possible behavior of that particular algorithm given the limitations of the demonstration data provided to it; moreover, given appropriate application and interpretation of such methods, quantification of absolute fiber connectivity can be achieved in this phantom, thus refuting the basis for their dismissal of such approaches for improving tractography quantification. The top row, left column of Figure 1A is a replication of both the Fiber Orientation Distribution (FOD) field and reconstructed streamlines from the aforementioned “toy example” phantom, which we have labeled “Patient.” In the top row, right column of Figure 1A, we have duplicated this phantom without incorporation of the simulated pathology or imbalance in streamline count between the 2 bundles; this we have labeled “Control.” Sarwar and colleagues describe the application of the SIFT method to the “Patient” phantom only. They correctly identify the fact that the SIFT method removes 2 streamlines from the lower unaffected bundle to match the tractogram to the underlying FODs (Figure 1A, bottom‐left). They then, however, erroneously conclude that: “Therefore, if the SIFT‐adjusted streamline count for Bundle 2 is compared between groups with and without pathology in Bundle 1, a spurious reduction in the streamline count will be found in Bundle 2 when in fact the true pathology is circumscribed to Bundle 1.” Comparing this result with application of the SIFT method to the “Control” phantom (Figure 1A, bottom‐right), it is clear that comparing absolute streamlines counts within a specific bundle across subjects, when the total number of streamlines in the tractograms of those subjects varies markedly for methodological reasons, would lead to erroneous inferences only because doing so would be inappropriate. Robust representation of this phantom and reconstruction using SIFT requires the following 2 modifications: