ROI-based tracks AFD, FA, and MD


#1

Hello experts,

I have generated tracks connecting two manually delineated ROIs in a group of subjects (disease and control). I’d like to assess how individual differences in participants’ track integrity (track weighted FA / MD, and fiber density, or AFD) relate to behavioral measures. I would really appreciate your advise with regards to the pipeline I’m planning to use to pull those numbers.

  1. Track-weighted FA / MD. I have done a full brain 10M tck and ran SIFT2 on those for each subject. I have created whole brain track weighted FA and MD maps using my “sifted” 10M tck. Now, I have generated my ROI-based tracks via tckgen, selecting equal number of streamlines connecting two ROIs in each subject. I was wondering if I should convert this ROI-based track into a simple ROI (binarize it? can I do it with MTRrix based commands (I could not find anything suitable)), so I can pull the FA / MD numbers for this track of interest from the whole brain track weighted FA / MD? Alternatively, I could go the simpler route and just run tcksample using my ROI-based tracks of interest and FA / MD maps that were not track weighted with whole brain tck. Which method would you recommend and why?

  2. Since I am not planning to do voxel-based group comparisons, do I still need to work with normalized FODs (and group RF to compute FODs) to build both whole brain tractogram and ROI-based tracks? Both 10M tracks and ROI-based ones were done on non-normalized individual RF FODs. After re-reading original Raffelt 2012 paper, I recon that for the track-weighted FA / MD this probably will not matter much, but for the fiber density measure I want to obtain next I am in trouble :slight_smile: , even being interested in individual differences. Would like to hear your thoughts on this, please!

  3. Now, the most interesting question. In order to get fiber density for my ROI-based track of interest, I have used tck2fixel (after having computed fod2fixel for the whole brain on NON-normalized individual RF FODs), and then mrstats to get the numbers. However, when comparing to other literature using MRTrix, my numbers for the mean pulled by this method (~0.4) are not up the par, and medians are 0. I wonder if this is because it doesn’t “remove” the rest of the brain’s fixels when selects the fixels for the track of interest, but kind of “zeros” them out? (sorry if this is completely incorrect understanding, I would really appreciate your explanation!). I found afdconnectivity command, and I wonder if this is the right way to go to compute AFD for my ROI-based tracks of interest? To make sure I understand correctly: in order to use wbft function, I need to generate my ROI-based tracks using tckedit command, correct? I have played with it a bit, and the number of streamlines generated for each subject with tckedit are very different. I wonder if this is because I was using individual RF non-normalized FODs -based whole brain tracks? If afdconnectivity is not the right command, could you please advise the correct route to compute this?

Thanks much for all your support and for this great tool!
Olga


#2

Hi Olga,

I was wondering if I should convert this ROI-based track into a simple ROI (binarize it? can I do it with MRtrix based commands (I could not find anything suitable))

It’s just a matter of decomposing the process into the underlying steps. tckmap can give you a map of the number of streamlines traversing each voxel; mrtransform can take an image containing such values and apply a threshold to produce a binary mask.

… so I can pull the FA / MD numbers for this track of interest from the whole brain track weighted FA / MD? Alternatively, I could go the simpler route and just run tcksample using my ROI-based tracks of interest and FA / MD maps that were not track weighted with whole brain tck. Which method would you recommend and why?

If you are intending to sample values underlying each streamline along its length, why generate a track-weighted image first? It would be a rather unusual quantity to be interested in, taking the “track-weighted-image-value underneath each track”.

Since I am not planning to do voxel-based group comparisons, do I still need to work with normalized FODs (and group RF to compute FODs) to build both whole brain tractogram and ROI-based tracks?

To be clear, by “normalized FODs” are you referring to the use of intensity normalisation? Spatially normalized FODs from individual subjects should not generally be used for tractography.

Intensity normalisation is a requirement for comparison of absolute values of AFD between subjects. If you are not doing this in your experiment, then it leads that it is not a requirement for your experiment. Measures like FA and MD are intrinsically ‘normalised’ within the diffusion tensor model.

In order to get fiber density for my ROI-based track of interest, I have used tck2fixel (after having computed fod2fixel for the whole brain on NON-normalized individual RF FODs), and then mrstats to get the numbers.

Just as important as which commands were used is what information they were provided with. It’s not clear to me exactly what is being quantified here. I think what you’re trying to do here is calculate the mean fibre density within a fixel mask, where that fixel mask is defined using a combination of tck2fixel and thresholding. If this is the case:

However, when comparing to other literature using MRTrix, my numbers for the mean pulled by this method (~0.4) are not up the par, and medians are 0. I wonder if this is because it doesn’t “remove” the rest of the brain’s fixels when selects the fixels for the track of interest, but kind of “zeros” them out?

mrstats will by default consider all values in an input image when computing statistics. If the input is a fixel data file, this means that all fixels will be included in the calculations. If you expressly do not want this to be the case, you can either use either the -mask or -ignorezero option.

I found afdconnectivity command, and I wonder if this is the right way to go to compute AFD for my ROI-based tracks of interest?

This is one way to go about deriving a quantity that is related to AFD. The quantity it provides is quite different to what is described above. The latter aims to quantify the mean fibre volume per voxel. afdconnectivity tries to provide a value that is proportional to total fibre cross-sectional area. So while they both rely on a fixel-wise measure of AFD, how the values across fixels corresponding to a particular pathway of interest are accumulated differs, and the fundamental units of the results differ also.

To make sure I understand correctly: in order to use wbft function, I need to generate my ROI-based tracks using tckedit command, correct?

Yes, as described in the documentation, it is a requirement of the -wbft option that the input tracks of interest be a subset of the whole-brain tractogram provided via the -wbft option, which means selecting them from the tractogram using tckedit rather than generating new tracks using tckgen.

I have played with it a bit, and the number of streamlines generated for each subject with tckedit are very different. I wonder if this is because I was using individual RF non-normalized FODs -based whole brain tracks?

The variance in streamlines tractography extends well beyond the use of individual response functions. Extracting a specific pathway of interest using tractography will demonstrate a large amount of variance, which is due to genuine differences in connectivity strength, differences in the manifestation of streamlines tractography biases, and the incredible difficulty of reconstructing trajectories through space based on noisy low-resolution estimates of local orientations. But both of the quantification approaches here (mrstats within a fixel mask, and afdconnectivity) specifically do not scale linearly with the number of streamlines, precisely for this reason.

Rob