Whole-brain tractography including subcortical structures

Dear MRtrix expert,

I read (on the batman tutorial) that using GMWM interface as seeding mechanism tend to penalize subcortical connections. A small pilot I did on my data seem to confirm this. As suggested from the tutorial I was going to concatenate the whole-brain tractography ( -seed_gmwmi) with the results of a second tractography in which I employ -seed_image with a mask of subcortical structures (basal ganglia, amygdala, hippocampus and “midbrain”) as input. However, I am not sure if the sifts2 framework would still be applicable in this context. The data are not formidable to start with (b-values 0 and 1000; directions: 30 or 48 in a subset of cases; nevertheless I would like to apply sift2).

An alternative would be to use dynamic seeding. The purpose is to generate connectivity matrices. Would you suggest the former or the latter approach?

Another related question is about combining cortical and subcortical regions in whole-brain tractography. I know that this a bit of a speculation but my understanding is that not everybody agree on dealing with cortical and subcortical structures in the same way. In my lab we encounter the problem using deterministic tractography with DTI-quality data, where it appears that subcortical structures “steal” streamline from cortica-cortical connections especially in cases presenting frank atrophy (cause by neurodegeneration). In a small pilot I run using MRtrix3 (taking a probabilistic approach, FOD2) on the same data the problem do not appear to be evident, but I was wondering if you would have any comment on the topic.

Many thanks in advance!

Nicola

Hi Nicola,

Interestingly you can actually combine -seed_gmwmi and -seed_image in a single tckgen call. The relative number of seeds drawn in each would be dependent on the relative volumes of the two seed images, which may or may not be “ideal” in any sense. Also, using the default output of 5tt2gmwmi as the input to -seed_gmwmi will actually seed streamlines “at” the sub-cortical grey matter structures, just at their interfaces with white matter rather than inside of them. But for whole-brain tractography intended for SIFT2 I don’t know that there is any situation in which I would advise not using dynamic seeding and instead using a less data-driven alternative.

The condition for SIFT2 to be applicable in this context is that the input tractogram needs to be whole-brain. If the particular type of seeding used tends to over-reconstruct or under-reconstruct certain pathways, that’s fine, that’s exactly what SIFT2 aims to address. But where it starts to break down is if pathways are wholly absent from the tractogram, whether due to the inability to reconstruct such or erroneous removal of such streamlines by the user prior to SIFT2.

As far as your final point, I’m not sure whether or not my interpretation is correct, so I’ll spell that out in addition to my response. What it sounds like is that, instead of a streamline traversing from cortical region A to cortical region B, the streamline instead (perhaps erroneously) intersects sub-cortical region C and thus terminates, resulting in a connection from A to C. I’ve had this problem myself reconstructing the CST when using ACT: over-estimation of the spatial extent of sub-cortical grey matter structures, combined with the ‘wiggliness’ of iFOD2 with the default power parameter, makes it rare for streamlines to traverse the entire internal capsule unimpeded. It’s true that in some instances this may be a combined false positive & false negative, especially if talking about tensor-based tractography. An alternative consideration is that it’s entirely possible that when performing whole-brain tractography, another streamline will be produced from cortical region B to the same part of sub-cortical region C, such that the connection from A to B is sort of present, it’s just via a second-order connection through C. There’s scope for alteration of analyses to take such data into account in a sensible way, but it might be too tangential given your actual question.

Cheers
Rob

Hi Rob,
thank you very much for your reply! I will move forward with using dynamic seeding.
I apologize, my last point was not clear but you got it right in any case :slight_smile:

If you don’t mind to elaborate more on this point it would be tremendously useful! I do not have in mind a specific tract because the “end point” of the analysis should be a structural connectome, which, of course, still entails some sort of “approximation” of the ““real”” pathways that connects each (sub)cortical region.
Thanks again,
Nicola

It’s tough to summarise in a succinct way, since whether or not the ideas translate to a particular experiment depend on the details of that experiment. But I’ll have a quick go anyway.

Let’s use the previous description: One expects a connection between regions A and B, but to a greater or lesser extent some streamlines (erroneously or otherwise) terminate in region C positioned in between them, such that instead of streamlines connecting directly from A to B, there are streamlines from A to C and streamlines from B to C. It would be reasonable to say that regions A and B are not entirely disconnected, since information could theoretically flow between A and B via C (and indeed this is the reality of some sub-cortical structural connectivity).

So one could then pose an open question: how do you “characterise” the connectivity between A and B, if you do not want these “indirect” connections to be inconsequential?

There’s a huge scope of possibilities there, but I’ll mention just a couple:

  • In graph theory one can consider the capacity for information transfer between two vertices taking into account all possible paths, not just the shortest path; this is referred to as communicability.

  • Heuristically, you could consider doing something similar but, instead of relying on a macro-scale parcellation, utilising the distances between streamline endpoints to modulate how much the communicability “decays” in the process of getting from A to B via indirect paths. So if there’s two streamlines, one from A to C and one from B to C, but their terminations in the thalamus are really close to one another, it could somehow contribute to your quantification of connectivity of the pathway between A and B not much less than would a direct connection.
    (Yes, there’s all sorts of assumptions and fudges in here; I’m mostly spitballing; but if anyone’s aware of where someone has done something like this, or pursues it themselves, do let me know)

Rob

Hi Rob,
thank you very much! I will keep thinking and piloting around this topic and I will let you know if I come up with (at least potentially) interesting :slight_smile:
Best,
Nicola