Virtual lesions on IIT atlas?

Hi all,

My knowledge of MRtrix seem to be a bit rusty (based on v0.2).

I want to achieve connectome tractography (pairwise streamline values, I think) using the IIT HARDI atlas. However, I want to apply virtual lesions to it, such that streamlines passing through the lesion area should be eliminated before computing the connectome matrix. I also need to do this for several virtual lesions.

What is the best strategy to use here? Run global tractography -> eliminate streamlines -> run connectome2tck? Or use the ACT pipeline with a fifth lesion tissue (this confuses me a bit). Or shall I just run pairwise tractography between parcels and then loop to check the numbers after streamline elimination?

What command can I use to eliminate streamlines in any case?

Note also that my intent is to use the Shen 268 parcellation. So another point of confusion to me is whether I need to run global tractography from the GM/WM border, or just use pairwise parcels as ROIs…

Thanks for any help.

Hi,

Firstly, I’m going to go on the assumption that you’ve made your mind up regarding applying virtual lesions to the tractogram, and assessing the subsequent affect on the connectome. I think the majority of research in this area applies virtual lesions to the connectome matrix directly; which of these is appropriate will depend on the particulars of your experiment.

Given you are looking at connectome generation, I would probably side with streamlines tractography using ACT rather than global tractography (though I may be ever-so-slightly biased…). I only know of one form of global tractography that provides any guarantees regarding the track termination points being in GM; others will commonly have many tracks that don’t reach the GM, and therefore will not form a part of your connectome, irrespective of lesion introduction. This means that although the tractogram matches the image data, the ‘tractogram-contributing-to-the-connectome’ may not.

Regarding streamline elimination, this will most likely be done using the tckedit command. The help page for that command should give the required details. I’d probably shy away from using the ACT fifth tissue type (or indeed playing around with the other tissue types) in this case: the tckedit approach will allow you to iteratively refine your ROIs and visualise which streamlines are and are not being selected, and will also be much easier to justify in a manuscript.

For connectome formation I would definitely use tck2connectome rather than using pairwise parcels as ROIs. The mechanism for assigning streamlines to GM nodes in that command is specifically designed so that the process behaves as one might intuitively expect:

  • If using pairwise parcels as ROIs, it becomes possible for a single streamline to traverse more than two ROIs and therefore contribute to more than one edge in the connectome.
  • If a streamline terminates precisely at the GM/WM interface, it may not quite penetrate into a voxel within the parcel. tck2connectome will account for this.

Though I did notice that the GM delineation is relatively narrow in that parcellation despite being a volumetric atlas: this may result in streamlines terminating at bottom-of-sulci not reaching a parcel. You may want to experiment with the value of the -assignment_radial_search option in tck2connectome. The connectome2tck command can also be used to investigate those streamlines not successfully assigned to a node pair.

Also: Global tractography cannot be ‘run’ from the GM/WM interface: individual tracks don’t have seed points as does streamlines tractography. Instead, track ‘segments’ are initialized throughout the WM, and these progressively connect to one another.

Hope that all makes sense!
Rob

Hi Rob,

Thanks a lot for the thorough response. After reading a bit of papers and documentation I definitely thought ACT is the way to go. But I wanted to ask some more questions, if you don’t mind.

  1. ACT seem to benefit maximally by the anatomical shape of the brain. But the IIT FOD atlas is an average of 72 people, the shape of the brain is also averaged (with the T1 being also blurry). Is ACT still useful in this context? I would think so but I’d appreciate your opinion.

  2. To make sure, the act file (i.e. 5tt.mif) contains probability images, not binarized segmentations, right?

  3. Is there a structurally based parcellation I could use? I could not find one from a quick look of the literature. And since I am using the Shen parcellation for some functional data, I though’t might be good to match them. The Desikan/Destrieux atlases that come with IIT atlas have large parcels including the entire temporal lobe gyri, clearly not well suited to distinuish specific tracts ending in perisylvian areas.

  4. I plan to bring the Shen atlas in IIT atlas space, and I don’t expect it to match perfectly the gray matter of the IIT atlas. To resolve the issue, I am thinking perhaps to expand the parcels in ANTsR until all gray matter of IIT is covered. Then remove voxels that do not fall in gray matter, to allow a good ACT run. But you mentioned the “-assignment_radial_search” which may be useful in this context. Do you have a feeling if one strategy might be more accurate than the other?

  5. I tried to find a solution by lesioning the connectome matrix, or by using lesioning tract probability maps of IIT, but I can’t come to a good solution. In principle a lesion can disrupt A-to-B connection at 80% and leave the remaining streamlines intact (I have real patient lesions). This information cannot be computed easily from a matrix or from voxelwise probability maps, unless I really check which streamlines does the lesion “cut”. I am happy to hear ideas on how to do this without going through tck* commands though. Unfortunately the .tck files are not available with the IIT atlas.

1.ACT seem to benefit maximally by the anatomical shape of the brain. But the IIT FOD atlas is an average of 72 people, the shape of the brain is also averaged (with the T1 being also blurry). Is ACT still useful in this context? I would think so but I’d appreciate your opinion.

Ah OK, I completely missed that you said you were going to perform tracking on the HARDI atlas. In our own work we haven’t used ACT when tracking on population template images, but we’ve been focused on fixel measures rather than connectome properties in this context. It may still be advantageous in a number of ways, particularly in ensuring that streamlines terminate in appropriate locations, but the magnitude of that effect may be less than that observed in single-subject data. The tracks themselves may however appear very ‘cropped’ and ugly-ish compared to non-ACT streamlines as they don’t project into the cortex at all. One trick you can try is taking the cortical GM image, adding it to the sub-cortical GM image, then zeroing the GM image: This will make ACT apply the sub-cortical GM priors in the cortex, allowing tracks to project into the cortex (but not allowing them to re-enter WM).

2.To make sure, the act file (i.e. 5tt.mif) contains probability images, not binarized segmentations, right?

Yep, it can make use of partial volume images. It will terminate streamlines once they cross the sub-voxel location where the tri-linear-interpolated value changes from ‘more-WM-than-GM’ to ‘more-GM-than’WM’. That way, you don’t get funny voxel-shaped jagged edges where the streamlines terminate.

3.Is there a structurally based parcellation I could use? I could not find one from a quick look of the literature. And since I am using the Shen parcellation for some functional data, I though’t might be good to match them. The Desikan/Destrieux atlases that come with IIT atlas have large parcels including the entire temporal lobe gyri, clearly not well suited to distinuish specific tracts ending in perisylvian areas.

This comes back to my misunderstanding at 1. When volume-based parcellations are registered to per-subject images, subject variation can have a significant influence on the overlap between the parcels and the subject’s cortex; this is why e.g. in the AAL atlas the parcels are massive. Whereas a surface-based such as FreeSurfer provides a parcellation image that is tailored to that subject’s cortical shape. Here, I would try using your parcellation as-is on the group template, but use the connectome2tck command to take a look at those streamlines where at least one endpoint wasn’t assigned to a parcel; that should highlight any problem areas.

4.I plan to bring the Shen atlas in IIT atlas space, and I don’t expect it to match perfectly the gray matter of the IIT atlas. To resolve the issue, I am thinking perhaps to expand the parcels in ANTsR until all gray matter of IIT is covered. Then remove voxels that do not fall in gray matter, to allow a good ACT run. But you mentioned the “-assignment_radial_search” which may be useful in this context. Do you have a feeling if one strategy might be more accurate than the other?

Personally I prefer the radial search mechanism to the parcel dilation approach. With the latter, there will inevitably be ambiguity about how to label voxels where two parcels intersect, and this uncertainty will propagate with subsequent iterations. With the radial search, the closest parcel to the sub-voxel streamline termination point will be chosen - up to the maximal search radius. So to me it’s more predictable.

Also, you wouldn’t strictly need to worry about removing non-grey-matter voxels: recall that ACT and the streamline-to-node assignment are two separate mechanisms that operate independently. The only benefit of such an approach would be if you then subsequently used a small-radius search, and the erosion of non-GM voxels would prevent streamlines terminating in WM from being assigned to the nearest parcel; but given the whole approach was predicated on using ACT that shouldn’t be present in your data.

5.I tried to find a solution by lesioning the connectome matrix, or by using lesioning tract probability maps of IIT, but I can’t come to a good solution. In principle a lesion can disrupt A-to-B connection at 80% and leave the remaining streamlines intact (I have real patient lesions). This information cannot be computed easily from a matrix or from voxelwise probability maps, unless I really check which streamlines does the lesion “cut”. I am happy to hear ideas on how to do this without going through tck* commands though. Unfortunately the .tck files are not available with the IIT atlas.

If you’re looking to perform image-based lesion simulation, I don’t think you have a choice other than to perform some form of tractography in order to assess the influence of lesions on the tracking, and concomitantly the connectome. Even if you were to have tract probability maps for every connectome edge, then you still couldn’t really assess the influence of a lesion ROI on the connection density: it’s the influence of the lesion on the tract cross-sectional area that determines the disruption of connectivity, ie. the lesion doesn’t necessarily have to infiltrate the entire tract volume.

Thanks Rob,

Would I still need to run tcksift as in the connectome tutorial?
So, the pipeline I devise should be:

  1. tckgen -act
  2. tcksift (???)
  3. tckedit
  4. tck2conenctome

By structurally based parcellation I mean tractography-based. Is there a parcellation based on tractography that might be worth using instead of my Shen choice?

Thanks a lot to all of you for doing such a great job with this software.

If you were to not use SIFT, the relative connection densities of different pathways in the brain would not be faithful, even before introducing simulated pathologies. This inevitably has a concomitant effect on any network properties you subsequently derive (manuscript proving this hopefully published soon). So I would definitely recommend including it to improve the biological interpretability of your experiment.

I’m not familiar with any whole-brain tractography-based parcellation atlases. Generally, tractography-based parcellation is limited to parcellation of a target structure / region based on connectivity to the rest of the brain; parcellating the whole brain based on connectivity to the whole brain is quite technically difficult, and is still an ongoing area of research. Everyone in the field accepts that the choice of parcellation for connectomics is still quite arbitrary at this point, so I wouldn’t be too concerned with your choice; the number of nodes is possibly more important than the precise source of the parcellation.