Creating a Structural Network with a Volmetric atlas

Hello everyone,

For a research project investigating structural connectivity in a clinical population, I am using MRtrix for streamline estimation & creating an atlas-based. I am having difficulty with the steps of preparing the volumetric atlas (using MCALT, Mayo Clinic Adult Lifespan Atlas) and then using this atlas for creating the structural network.

The BATMAN tutorial, which is incredibly useful, they use a FreeSurfer-based atlas, whereas the MCALT is not, so this most likely necessitates other processing steps.

Based on previous entries into this community - am I correct in thinking that I need to

  1. register subject-space T1 to atlas
  2. invert the transformation to get the transformation matrix
  3. apply the reverse transformation to the atlas so it’s mapped on the subject T1
  4. generate the connectome in subject space

I have two questions surrounding this procedure:
Q1: In the first step, where you register T1 to the atlas, you would do this to the T1-image that is coregistered to the DW-image, instead of the raw T1 to have them in the same space?

Q2: Which processing software do you advise to do the MRI-atlas registration with? Does mrtransform (from MRtrix) suffice, or would advise me to use other software/functions, such as flirt from FSL or ANTs - taking into account the balance between efficiency and accuracy.

Hi @Bastiaan_van_Tol and welcome!


Work with the preproc T1 registered to your DWI.

I personally prefer ANTs, but do whatever works for you. Just be sure to do nearest neighbor interpolation when you warp the atlas.


Hi Steven,

Thanks a lot for your input, I appreciate it.

One follow-up question: I have executed the following processing steps

flirt -in T1_coreg.nii -ref MCALT_atlas.nii -out T1_to_atlas.nii -omat T1_to_atlas.mat -interp nearestneighbour -dof 12

convert_xfm -inverse T1_to_atlas.mat -omat atlas_to_T1.mat

flirt -in MCALT_atlas.nii -ref T1_coreg.nii -applyxfm -init atlas_to_T1.mat -out atlas_to_T1.nii -interp nearestneighbour -dof 12

But my output is unsatisfactory, as you can see in the image below!

I get the sense that you and others (as described here) that this can be common with FSL, and that it is generally harder to get everything right compared to ANTs. So, thanks again for your advice, I will try using ANTs and see how this performs.


Hi @Bastiaan_van_Tol,

It would be better to register a T1 from the atlas space (e.g., maybe it’s in MNI space) to your subject T1. The similar contrast will lead to a better registration.


Hi Steven,

Aha okay! That makes sense, thank you

So that means - the ‘pipeline’ would look like this:

  1. register atlas-space T1 to subject-space T1
  2. invert the transformation to get the transformation matrix I want to ouse in my next step
  3. apply the reverse transformation to the atlas image so it’s mapped on the subject-space T1
  4. generate the connectome

I hope I understand your point and solution correctly, thanks!


1 Like

Yup, that should do it!

Hi Steven,

I have tried your suggested steps ANTS with the following script:

ANTS 3 -m "PR[MCALT_T1_05mm.nii,T1_coreg.nii, 1, 2]" -o ANTS -r "Gauss[2,0]" -t "SyN[0.5]" -i 30x99x11 --use-Histogram-Matching

WarpImageMultiTransform 3 MCALT_ADIR122.nii MCALT_subjectspace.nii -R T1_coreg.nii -i ANTSAffine.txt ANTSInverseWarp.nii --use-NN

Now, the problem that occurs in this solution that the MCALT atlas does move outside the subject space cortical boundaries of the MR and DW images, shown below.

Do you have idea what could be the problem here? Would you suggest changing up some registration parameters to find what works best, or do you have specific thoughts in this specific use case?

Thanks a lot!

Can you show what you subject DWI and T1 look like coregistered, and the atlas T1 registered to your subject T1? And the atlas on the MNI template T1. If any of these are off, it could lead to misalignment.

Hi Steven,


subject DWI and subject T1 coregistered:

Atlas T1 registered to subject T1 (Here the problems seems to occur, as the pattern of the mask not fitting the cortex well looks highly similar as the DW-atlas problem):
Screenshot 2024-02-22 at 10.18.45

Atlas registered to the MNI template T1 (This I did not compute, as both the MCALT atlas and the MNI template in T1 were already in the same space when I downloaded the atlas, and it looks fine)
Screenshot 2024-02-22 at 10.20.55


That second image is the atlas itself overlaid the subject T1, I am interested in the atlas-space T1 (e.g., the MNI T1) co-registered to the subject space.

Hi Steven,

A late reply, but I have tried some processing alternatives.

Do you suggest skull-stripping both the subject T1 and template-space T1 before non-linear registration between them?

This is the result of subject T1 registered to template T1 (no skull stripping). You can see that the cortical boundaries in the overlayed T1 image (in red) seem to stretch further out compared to the T1 template (like stretched all the way, no space between brain tissue and meninges):

And this is the result of the subject-T1 to template-T1 registration when I have used skull-stripping. I overlayed the skull-stripped registered T1 to the regular template-T1 image so you can see it better (here there is space between brain tissue and meninges):

In this new pipeline, the subject DW registered to subject T1 looks like this (very good it seems like):

However, when applying the inverse transforms now to my atlas parcellation (and using -n GenericLabel / NearestNeighbor) - This still does not look entirely optimal, as some amount of cortical space in the DW-image escapes the atlas. Is this typical, and not entirely problematic (as good as it gets), or would you say this is going to negatively influence my tractography, and I need to find a more optimal solution?:

Thanks a ton,

Hi @Bastiaan_van_Tol,


You could try different registration algorithms between the T1s, but this registration looks pretty good now. Given how large these parcels are, I doubt minor optimizations are going to change anything meaningfully.

I also recommend reading this article which talks about considerations about parcellations in connectomes: (Yeh et al., 2019; NeuroImage)