More specifically, I guess that in
tck2connectome only streamlines starting from (some maximum distance from) and ending in (some maximum distance from) the Grey Matter (GM) will be retained and included in my connectivity matrices. Thus, I wonder whether seeding from the GMWMI in this case would be more appropriate than seeding from the White Matter (WM) mask, since by seeding from the GMWMI I guess that it would be more probable that the reconstructed streamlines get assigned to the parcellation nodes in the GM (both cortical and subcortical) as compared to possible streamlines starting in seeding voxels in the middle of the White Matter (WM) mask.
While I don’t have the relevant data at hand, I would not expect the proportion of generated streamlines successfully assigned to the connectome to vary massively between WM seeding and GMWMI seeding. If ACT is being used, then streamlines seeded within (what is segmented as) WM will not be any more likely to terminate within (what is segmented as) WM than streamlines seeded at the GMWMI, since it’s explicitly forbidden in either case.
What actually leads to streamlines being generated and accepted by ACT but not contributing to the connectome are:
Streamlines terminating in locations where biologically they should be permitted to terminate, and based on the provided tissue segmentation ACT does permit streamlines to terminate there, but no corresponding label is present within the particular parcellation scheme used;
Streamlines terminating in locations where biologically they shouldn’t be permitted to terminate, but ACT has been provided with a tissue segmentation image that results in it erroneously permitting streamlines to terminate there.
The probability of either of these situations occurring is, as far as I can predict, no different between seeding homogeneously in (what is segmented as) WM v.s. seeding at (what is segmented as) the GM-WM interface. If anything the latter may be fractionally elevated if seeding at the GM-WM interface (based on the mechanism by which such seeds are distributed), which would be contrary to your own a priori expectation. It’s an experiment that you already have the requisite tools to perform if you want to know for sure.
Which seeding strategy would you recommend for a connectomics analysis?
The more reconstruction density biases are present in the initial tractogram, the more work SIFT2 has to do to correct them. As shown in the SIFT2 manuscript, the dynamic seeding approach provides a better initial solution than does WM or GMWMI seeding; as one would expect given that it’s applying the same model and attempting to drive toward the same solution as does SIFT(2).
Does the seeding strategy matter after the application of SIFT2?
Ideally, the resulting connectome following SIFT2 would be independent of the particular seeding strategy used to produce the tractogram prior to SIFT2; it won’t be exactly the case due to e.g. regularisation, but it should at least be more similar than if SIFT2 had not been used. I’ve shown this for SIFT in a fairly crude manner, and further evidence was produced here by @chunhungyeh.
Which are the mechanisms used by the
-assignment_forward_search flag in
tck2connectome command in order to assign streamlines to the closest cortical / subcortical parcellation node?
The “forward search” takes the streamline endpoint and the tangent at that location, and extends a ray from the endpoint in that orientation, waiting to intersect a voxel that has a label ascribed to it. In part the implementation for such is there (along with others) in order to demonstrate the fact that there must be some explicit process by which streamlines are assigned to parcels, that there is no single unambiguous choice regarding how this should be done, and that there is most definitely scope for resulting connectomes to vary depending on the mechanism used. This was later described here; the “forward search” mechanism itself wasn’t described there, but the fundamental point stands.
Is there any recommendation about whether to choose
The “radial search” is the default mechanism that is utilised if no explicit choice is made, so is the “recommendation” as it were. My concern with the “forward search” is that if a streamline intersects the GMWMI at close to perpendicular to that surface, then a ray projected from that termination point in the direction of the final streamline tangent may not intersect an appropriately labelled GM voxel. After having mucked around a bit with this problem at the time of creating
tck2connectome, I ended up being content with the radial search for the time being. I have what I think is a better mechanism in private code, which uses the ACT 5TT image to better inform the direction in which to search for a label, but that’s bundled with a suite of other changes that I’ve never found the time to resolve. Using native surface information directly (as long as the same surface is used for both tracking and parcellation assignment) resolves the ambiguity around streamline-to-node assignment, but I don’t know how far away @chunhungyeh’s code might be from inclusion in MRtrix3.