Registration to atlas or atlas registered

I would like to do something like voxel based analysis within basal ganglia with the AAL3.
Now, would you register all subjects to the atlas or its referent image (like in the old good VBM pipelines)? or register the referent template and then with the computed matrix warp the atlas into each individual T1?


The former: you should register each subject brain to the template space, which is MNI for the AAL.

Why only the former?
If the registration is linear or (what its most common) an elastic diffeomorphic algorithm (such as SyN of ANTs), it should not matter?

Shouldnt the decision be done based on template vs original image resolution?

I say the former because voxel based studies usually rely on having all subjects in the same space.

Hi, this works for me as well. When registering the diffusion images to MNI space, one should also correct the B-bevcs or no? I mean rotate the bvecs (as it is done after motion correction). I think it is reasonable, but I have failed to find any confirmation of this step. Thanks.


Yes, any registration of DWI images should be accompanied by rotating B-vecs.


Hi! Thanks for the reply!
I think so too, but discussing this with a colleague he mentioned an interesting point, which is that after a rotation, the angles of the bvecs with respect to the system’s axes havent changed so the bvecs coordinates should remain the same. I am not entirely sure about this argument but he made me doubt. Is there a reference or paper to cite when doing this preprocessing step? It is true that a most of the works register the atlas to the diffusion space, but there should be also the other case. I guess we are just looking for a confirmation on this matter. Once again, thanks again for the prompt answer.

The important thing is that the bvecs orientation has changed relative to the brain’s axes. To illustrate this point, consider an extreme rotation that rotates the brain 90 degrees (such that the x-axis becomes y-axis). Let’s say that originally (pre-rotation), the gradient was pointed in the pure +X direction, and this image indicated a high degree of diffusivity in a particular voxel. If the brain is now rotated, and the gradient not rotated, that volume would still indicate a high diffusivity in the X direction relative to the scanner, but in the brain it would appear the diffusivity is now going in the Y direction, which can completely confound tract tracing efforts. If your goal is only to produce scalar maps (e.g. FA, MD), then maybe you do not need to rotate, but if you need directionality for any purpose, then you will need to account for this rotation in your gradients. I do not have a citation off the top of my head that supports this, but it is part of tract segmentation workflows such as TractSeg.


Hi Steven!
The confirmation I was looking for… because it’s exactly what I was thinking but expressed in better terms. Thank you very much for the very detailed and educative answer!!