Population_template from diffusion tensor or kurtosis tensor?

Hi all,

I was wondering whether the population_template command also works to make a diffusion tensor template or kurtosis tensor template?

A first attempt using diffusion tensors as input failed, but perhaps there is a workaround using the options?



Hello Thibo,

We don’t have support to register and reorient tensor images. An alternative would be to derive scalar images from the tensor image using tensor2metric and register those.

To mimic diffusion tensor, you could also use spherical harmonics of order lmax=2 as the input to population_template:
dwi2fod csd input.mif wm_response.txt fod.mif -mask brainmask.nii -lmax 2.


…although nothing would stop you from doing CSD up to any lmax that your data supports to build the best possible template (and in that way get the best warps to template space). You could then still use those warps to warp the maps (certain DKI metrics?) you’re looking to use into a common template space. But it’s indeed not possible to correctly warp and reorient tensors with MRtrix (and hence use the full tensors themselves in template space).

Alright, thanks Max and thanks Thijs. The goal was indeed to warp DTI parameter maps to a population template using optimal warps. The suggestion to use FODs with optimal lmax will be perfect, thanks!

1 Like

Building on this:

When using FODs for registration purposes only, would it be ok to:

  1. Skip the dwi intensity normalisation step, instead use the preprocessed and bias field corrected dwi.
  2. Skip the averaging response functions step, instead use the subject specific response function for FOD estimation.

I guess both intensity normalisation and use of the average response function are necessary for fixel based analysis. If the FOD would only be used to create a template to which some scalar maps are to be warped, perhaps these steps can be skipped?

Yes, that would be fine. FOD registration uses a mean squared metric, so the image intensities need to be in the same range, but using a subject specific response for CSD will help with this.

Agreed! The funny thing is that both of your proposed “skips” should actually work very well together but not separately; i.e., it would rather be a problem if you did one of those skips but not the other. Doing them together effectively lets the response function selection algorithm act as a normalising mechanism, that should render the FODs (and even multi-tissue contrasts) as similar as possible… which is effectively what you need for registration (and/or template building). In all our discussions on normalisation, we’ve even successfully used the response function selection algorithm(s) (partially) for this purpose at some point.

What you can use as well or in addition for your scenario, if you’d like to give it a shot, is checking out the tag_0.3.16 branch of Mrtrix, and using the new mtbin command on your multi-tissue CSD output. Even better for this particular purpose (where you will not be using the FODs and tissue types for further analysis, but you’re just looking for the best performance on the registration/template front), would be to use the mtbin command with the -independent option.

Thanks a lot both for the explanation!