Fixel-based analyses of laterality

Hello, MRtrix experts

I was doing a FBA of laterality following the pipeline in this paper : Helena Verhelst, Thijs Dhollander, Robin Gerrits, Guy Vingerhoets,Fibre-specific laterality of white matter in left and right language dominant people,NeuroImage (2021)

But I got stuck by several questions.

  1. How to make a single hemisphere fixel mask. I used to make masks based on atlas, but I had no idea how to produce a mask in fixel format based on the study-specific template. And in the paper, they referred that "restrict the fixel analysis mask to only contain fixels for which correspondence [fixel mask and those of individual subjects] with at least 90% of all participant hemisphere data could be established. I wondered how to accomplish this step and combined with hemisphere mask .

  2. In order to calculate log-ratio of laterality index [log(FDoriginal/FDfliped)], in this paper they set the FD to a non-zero low value of 0.01 instead (this equals 1% of one calibrated unit of FD) . I did this using command [ mrcalc fd.mif 0.01 -add fadadd.mif ] for each subject’s fd file, but I don’t know this step should be done before or after fd smooth.

  3. fixelcfestats seems not include One-sample T test. Are there other ways I can do this?

Thank you very much !

Best,

Kitty Zhu

Hi Kitty,

How to make a single hemisphere fixel mask.

Once you have your FOD template, and have created a voxel mask corresponding to one hemisphere, you can then generate a fixel mask containing 1 for all fixels within the selected voxels using the voxel2fixel command.

restrict the fixel analysis mask to only contain fixels for which correspondence [fixel mask and those of individual subjects] with at least 90% of all participant hemisphere data could be established.

This is just a matter of decomposing into the requisite components:

  1. Classify every fixel in every participant as either having established correspondence (FD > 0) or not (FD = 0) (e.g. mrthreshold -abs 0.0 -comparison gt)

  2. Quantify the fraction of subjects for which correspondence was established. If the output of step 1 is a binary fixel data file, containing 1 for correspondence and 0 for not, this is just the sum of those images divided by the number of participants. (e.g. mrmath to get sum across participants, then mrcalc to do the division)

  3. Apply a threshold of 90% (e.g. mrthreshold -abs 0.9)

  4. Take whatever fixel analysis mask you are currently using, and multiply it with the output of step 3 to produce a new fixel analysis mask that excludes those fixels with insufficient fixel correspondence.

set the FD to a non-zero low value of 0.01 instead (this equals 1% of one calibrated unit of FD)

  1. This is not an addition, but a max operation. You want to take the maximum of 0.01 and the empirical FD; that way, anything less than 0.01 (including 0.0) is replaced with 0.01, but any valid FD value above 0.01 is retained.

  2. I think the problem here is still slightly ambiguous, as it is not just the order of the max operation relative to smoothing that is unknown, but also the data that are actually to be smoothed. If the FD data were smoothed prior to computation of the log ratio, then this would remove zero FD values, and the issue of non-finite values would I think have been described differently by the authors to how it was. If it is the log-ratio that is to be assessed for statistical significance, it would be more “standard” to compute that measure for each fixel, and then apply smoothing to that measure. In this case it’s the raw FD values that would need to be clamped in order to ensure that the log-ratio values are valid everywhere prior to smoothing. You could ask the authors the precise sequence of steps they performed.

  3. I have an alternative method for establishing fixel correspondence that I’d like to get incorporated into version 3.1. The problem introduced by fixels with zero FD as described here was one of many motivating factors for that work. The theory being that the threshold(s) applied during fixel segmentation would be relaxed, facilitating fixel correspondence even for small fixels, and there would then be greater dependence on excluding those fixels that are small / inconsistent across subjects as part of derivation of a fixel analysis mask. Your data would potentially serve as a useful case study for revising advice on pipeline parameters, if you would be interested in engaging with that higher level of complexity.

fixelcfestats seems not include One-sample T test.

-errors ise. Please see GLM manuscript for equivalence between this shuffling method and a conventional one-sample t-test.

Cheers
Rob