FA template from FOD template?

Hello MRtrix community,

I have a dataset that was processed according to the pipeline described here:

I want to generate a study-specific FA template, but I only see mention of an FA template in the single-tissue CSD pipeline documentation, via the command “dwinormalise group”. Since I followed the multi-tissue pipeline, the data were normalized using the “mtnormalise” command (which normalizes FOD images), and I do not have an FA template output.

Is it possible for me to generate an FA template from the population template WM_FOD image?

Thanks for any advice!


Not directly. If you are after a measure of anisotropy and you don’t need FA, you can use the command sh2power -spectrum to calculate the power spectrum of the template FOD. The l=2 term (2nd volume) is conceptually comparable to FA.

If you need FA and if you saved the warps (-warp_dir) you can apply these to FA images generated from the dMRI data (dwi2tensor, tensor2metric) that were used to generate the FODs via mrtransform -warp_full and average the transformed images using mrmath (if not all images cover the full area of interest, after transformation you’d need to nan-mask areas outside the transformed masks using mrcalc.

If you don’t have warps, you can re-register the subjects’ FODs to the template (as in the FBA pipeline) and use these warps to transform the subjects’ FA images to template space but the two templates wouldn’t be 100% comparable as registration procedures inside population_template differ from directly registering a subject’s FOD to the final template.

Another option is to recreate the template using FOD and FA contrasts which would generate the template for you. You could set the multi-contrast weights for FA (-mc_weight) to zero if you do not want FA to contribute to the registration.

1 Like

Thanks for your response @maxpietsch ! My primary goal is not actually to measure anisotropy, but rather to use the FA template for this population as registration target from the MNI FA template brain, in order to have ROIs from MNI space transformed into the “template space” of the present study. The FOD template does not register between the standard T1 weighted MNI brain.

I’m thankful for any advice or comments you may have. I do have the warps saved, so I can generate each subject’s FA map in subject space and transform them (individually) to template space. Is mrmath able to average them in a meaningful way to represent a “template”, or would it be better to use something like ANTs to make a template derived from the individual subject FA images warped to template space?

1 Like

Hi Bradley,

Firstly, it’s important to distinguish between spatial normalisation, which one would achieve using e.g. population_template, and intensity normalisation, which is performed by dwinormalise group / mtnormalise. Your initial question might be conflating the two, so I want to make sure this is understood.

(What makes it even more confusing is that the dwinormalise group algorithm, which performs intensity normalisation, internally uses spatial normalisation in order to achieve that goal :exploding_head: )

Otherwise, I agree with @maxpietsch that using the non-linear transformations derived from FOD-based registration, applying them to each subject’s FA image, and then averaging the results, is likely the best choice. If the principal purpose of the template FA image is to obtain spatial correspondence with the MNI FA template in order to transform ROIs from MNI space into the space of your FOD template, then you do not want to be using any process that induces geometric differences between your FOD and FA templates. That means using the same subject-to-template spatial transformations for both forms of data.


1 Like

Thank you for the follow-up @rsmith ! You are correct that my end-goal with regarding to this post is spatial normalization, rather than intensity normalization.

I have followed this advice and now have a good way to get ROIs from MNI space to the FOD template space. Thank you!

1 Like

May I ask a similar question: how to generate a template for principal eigenvector maps? Would you suggest the same way as you described here about how to generate an FA template? Is it even possible to put only principal eigenvector maps volume in the input_dir instead of FOD volumes in the input directory?

Would it help the registration if I put b0 images and FA maps additional to the FOD volumes for multi-contrast registration?



how to generate a template for principal eigenvector maps?

One would not typically use such data in the construction of a template; it is not an appropriate contrast to be using to drive the registration process (consider e.g. the fact that two subjects, with eigenvectors [0 0 1] and [0 0 -1], would be considered to have very poor correspondence). Instead, you would want to perform template construction & transformation of data to the template, and then compute the principal eigenvector for each subject’s data within template space.

Since the principal eigenvector is a derivative of the diffusion tensor model, the most logical solution would be to use tensor-based registration and produce a tensor template image. While we don’t have this implemented in MRtrix3, using population_template for FOD-based registration with lmax=2 is very comparable. From there you could take the DWI data in subject space, apply the computed spatial transformations (without modulation) to get those data into template space, and then fit the diffusion tensor model and extract the principal eigenvector upon those data.

Would it help the registration if I put b0 images and FA maps additional to the FOD volumes for multi-contrast registration?

Maybe, maybe not. Given the use of a mean-squared image similarity metric, including b=0 volumes as-is as a second contrast would mean that differences in b=0 intensities would drive the registration much more so than FOD differences; and even if you scaled the image intensities in some way such that the FOD images and b=0 images contributed approximately equally, it’s likely that the use of b=0 images would prioritise the alignment of CSF rather than the WM. And while it’s tempting to say “more data = better”, including FA data also potentially introduces some downstream effects of the limitations of that metric, which could conceivably be detrimental. But this is not my area of expertise, and it’s possible that there are manuscripts out there exploring the utility of multi-contrast registration of which I am not aware.