SC matrices generation: is merged .tck the most appropriate option? And has ACT to be used when seeding from subcortical binary GM mask?

1

Dear experts,

I would like to produce structural connectivity (SC) matrices from some DWI data (5 b0, 100 volumes acquired with b=3000) with MRtrix. The nodes of the SC matrix that I would like to produce are both cortical GM and subcortical GM regions (e.g. Desikan atlas parcels).
I was using grey matter-white matter interface as seed for whole-brain tract generation. However, BATMAN tutorial reports:
“One drawback of starting the streamlines at the gray-matter/white-matter-boundary is that streamlines between subcortical regions are hardly reconstructed. Therefore, if you are interested between such connections, you should create an additional streamlines file where you seed from a (binary)subcortical mask only, and combine both results.”

As I am interested in including subcortical GM regions in my SC matrices, I changed my pipeline as follows:

tckgen -act 5tissue_pm_coreg.mif -backtrack -seed_gmwmi gmwmSeed_coreg.mif -select 10000000 wmfod_norm.mif tracks_10mio.tck

tckgen -act 5tissue_pm_coreg.mif -backtrack -seed_image subcort_gm_mask.mif -select 10000000 wmfod_norm.mif tracks_sub10mio.tck

tckedit tracks_10mio.tck tracks_sub10mio.tck merged.tck

tcksift2 -act 5tissue_pm_coreg.mif merged.tck wmfod_norm.mif sift2_weigths.csv

tck2connectome -symmetric -zero_diagonal -scale_invnodevol -tck_weights_in sift2_weigths.csv merged.tck aparc_parcels_coreg.mif aparc.csv -out_assignment assignments_aparc.csv

Here are my questions:

  1. Do you think that generating a whole brain SC matrix from merged.tck, in this specific case, is more appropriate than generating it from tracks_10mio.tck (i.e. the .tck file generated just with seed_gmwmi gmwmSeed_coreg.mif)?
  2. When using the subcortical gray matter binary mask as -seed_image for tckgen, has -act option to be included or not? In other words, for generating tracks_sub10mio.tck, which is the best option bertween the following?

a) tckgen -act 5tissue_pm_coreg.mif -backtrack -seed_image subcort_gm_mask.mif -select 10000000 wmfod_norm.mif tracks_sub10mio.tck
b) tckgen -backtrack -seed_image subcort_gm_mask.mif -select 10000000 wmfod_norm.mif tracks_sub10mio.tck

Can you please help me with solving these doubts, in this specific framework? I would like to avoid introducing biases in SC matrices due to metodological choices.
Best,

Laura

2

Welcome Laura!

One drawback of starting the streamlines at the gray-matter/white-matter-boundary is that streamlines between subcortical regions are hardly reconstructed.

This statement may be predicated on the GM-WM interface seeding image exclusively containing the cortical interface. If generated in such a way that the seeding image additionally contains the interfaces between sub-cortical grey matter structures and the adjacent white matter, you may find that the balance in reconstruction density between cortical and sub-cortical is reasonable. It’s been too long since I’ve looked at the raw data in this respect, as I would typically instead advocate the use of dynamic seeding across the board (and not only due to the software gymnastics I had to do to implement it!).

tckgen -act 5tissue_pm_coreg.mif -backtrack -seed_image subcort_gm_mask.mif -select 10000000 wmfod_norm.mif tracks_sub10mio.tck

If you were to go down this road, I would probably advocate (assuming that your sub-cortical grey matter segmentation includes partial volume fractions) that instead of binarizing a sub-cortical grey matter mask and using -seed_image, you could instead retain the partial volume fractions and use -seed_rejection.

I’ve thought at times about having a dedicated sub-cortical seeding mechanism that would provide a slightly better guarantee of homogeneous seed point distribution within these structures, but never got around to it…

Do you think that generating a whole brain SC matrix from merged.tck, in this specific case, is more appropriate than generating it from tracks_10mio.tck (i.e. the .tck file generated just with -seed_gmwmi gmwmSeed_coreg.mif)?

It depends on the magnitude of the reconstruction biases present in either reconstruction. What you’re ultimately looking for (ignoring other factors since you’re interested in whole-brain connectivity) is an initial whole-brain tractogram where the reconstruction density biases are minimal; that is, you want to minimise the “amount of work” that SIFT2 has to do in order to achieve correspondence between streamlines densities and fibre densities. You can actually test this directly: Use the -csv option, and see which reconstruction has the greater cost function prior to the first iteration. One of the reasons I generally advocate the use of dynamic seeding is because it yields a tractogram where the cost function prior to SIFT2 optimisation is less than that achieved using other seeding strategies (which is what dynamic seeding is designed to do, after all).

When using the subcortical gray matter binary mask as -seed_image for tckgen, has -act option to be included or not?

You would not want to be taking one tractogram generated with ACT, and another generated without ACT, and concatenate them.

In its original implementation, ACT had an issue with this kind of approach: seed points simply had to be classified as inside WM. However at some point I made it a little more clever. If a seed point is within sub-cortical GM, regardless of whether tracking is unidirectional or bidirectional, ACT should do “the right thing”: the streamline must exit the sub-cortical GM and enter WM in order to satisfy the anatomical priors, but in addition if tracking is bidirectional from the seed point then only one of the two propagations from the seed point can enter WM.

tckgen -backtrack -seed_image subcort_gm_mask.mif -select 10000000 wmfod_norm.mif tracks_sub10mio.tck

This wouldn’t work: -backtrack is exclusive to ACT.

I would like to avoid introducing biases in SC matrices due to methodological choices.

Similarly to previous: One of the goals of SIFT(2) is to reduce the dependence of estimated connectivity measures on the seeding strategy employed (see manuscript). After all, if there is a true underlying biological connection density, how could it possibly depend on tractography seeding? Changing seeding strategies can’t fix the reconstruction density bias problem entirely, but some strategies give a better initial estimate than others. I would generally suggest using dynamic seeding if interested in whole-brain connectivity; and that would pretty much absolve you of responsibility for making uniquely poor methodological choices in this regard.

Cheers
Rob

3

Thank you very much Rob for your explanation and suggestions, they are very useful. I’ll take your advice and go for dynamic seeding:

tckgen -act 5tissue_pm_nocoreg.mif -backtrack -seed_dynamic wmfod_norm.mif csffod_norm.mif -select 10000000 tracks_10mio_dynamic.tck

tcksift2 -act 5tissue_pm_coreg.mif tracks_10mio_dynamic.tck wmfod_norm.mif sift2_weigths.csv

I have just one last question: is it ok if I use wmfod_norm.mif and csffod_norm.mif as input for seed_dynamic option? Basing on my data (5 b0, 100 volumes acquired with b=3000) and on this previous discussion (Full processing pipeline with single phase-encoding direction in MS - #2 by rsmith), I used the multi-tissue CSD algorithm to derive FODs but provided only the WM and CSF response functions:
dwi2fod msmt_csd dwi_den_preproc.mif -mask brain_mask.mif -fslgrad bvecs.bvec bvals.bval wm_response.txt wmfod.mif csf_response.txt csffod.mif

I thank you in advance for your attention and help.
Best,

Laura