ACT Pipeline question

I have a couple quick questions with regards to my ACT pipeline. I have produced streamline count connectomes based on the AAL atlas which im confident of but now would like to check the last steps of my pipeline. These will be fed first into NBS for basic statistics (along with fMRI and MEG using the same ROIs) and then given to someone much smarter than me for machine learning purposes.

My original last step for the streamline count connectomes (based on the BATMAN tutorial and Andy’s Brain Book):

tck2connectome -symmetric -zero_diagonal -scale_invnodevol -tck_weights_in sift_1M.txt tracks_10M.tck parcels_coreg_AAL.mif parcels_coreg_AAL.csv -out_assignment assignments_coreg_AAL.csv

Now I want to produce connectomes using FA,MD,AD and RD. I have produced the scalar maps in FSL and upsampled them to 1.25 mm isotropic (I noticed now this isn’t necessary, but for now it’s for the sake of consistency). Based on the post here (FA, AD, RD connectomes) and the MRtrix documentation, I used:

tcksample tracks_10M.tck FA.mif mean_FA_per_streamline.csv -stat_tck mean

tck2connectome -symmetric -zero_diagonal -tck_weights_in sift_1M.txt tracks_10M.tck parcels_coreg_AAL.mif parcels_coreg_AAL_FA.csv -scale_file mean_FA_per_streamline.csv -stat_edge mean -out_assignment assignments_coreg_AAL_FA.csv

Essentially in the last step I removed -scale_invnodevo and added -scale_file and -stat_edge while keeping everything else the same. Numerically the connectomes make sense and visually they look accurate, but I just wanted to confirm as we are moving to a very large data cohort.

Lastly, in some of the the steps to bring the AAL ROI’s into diffusion space I noticed many tutorials work off only a linear registration. I added in a nonlinear registration to improve several steps (using nearest neighbour interpolation for applywarp) and the results look great. Is there any reason this step isn’t added in many tutorials beyond computation time and potential for problems?

Thank You

Hi Patrick,

Personally I’m not a big fan of the scaling by reciprocal of the node volumes; it’s there for feature-completeness, but it considerably alters the interpretation of the connectivity measures. I’ve probably ranted on it a few times on here, but I lose track of such things. The argument I give in the discussion of this preprint regarding intracranial volume extends to individual node volumes also.

The body of that preprint also does a better job of justifying the mechanism by which one can obtain estimates of white matter connectivity that should be better quantitatively than raw streamline count.

Yep that looks fine. Note that this may be different to how such quantifications are done elsewhere, as described in this wiki post.

Personally when I constructed my own pipeline I initially used affine registration for the same reason, but found the results to be entirely lacklustre and switched to nonlinear registration. So it is indeed possible that, for tutorials at least, people stick to linear registration for simplicity, but for an actual experiment I would myself recommend going nonlinear, even with a parcellation as crude as AAL.