From FSL, I derived the HarvardOxford probability images in FSL’s own taste of MNI space (MNI152NLin6Asym). I would like to have those in MNI152NLin2009cAsym space (not the maxprob image: I want a 4D image with one 3D image for every brain area included in the atlases, where the respective 3D images gives me the probability of a voxel being in this brain area).
As pointed out in this discussion before, templateflow archive nicely provides an ITK transform to map information from MNI152NLin6Asym space onto MNI152NLin2009cAsym. I applied this transform on the HarvardOxford-images from FSL using ants and got what I wanted, it looks nice on the first glance
(As a reference, I used the NLin2009cAsym-T1w-image provided by templateflow. Interpolation set to linear)
Now, just to see if everything checks out, I thresholded those pictures at 0, 25 and 50 and derived the maxprob-images. Then I compared those to the maxprob-images provided by templateflow. They are like 95% the same and therefore I assume I didn’t make a huge mistake like wrongly thresholding or applying the wrong transform or something like that. But at the edges of several brain areas, there are some differences between the two.
So since I am rather new to MRI imaging, I have to ask: Two maxprob-images of the same atlas in the same space differ in their information - is that, like, ok or normal? Would one expect differences there? Does anyone know where these differences could come from?
And most importantly: Is the transformed probability image still usable on fMRI data that is in MNI152NLin2009cAsym space?
I suppose you could find small-to-moderate differences depending on specifics of how the registration was calculated and applied. Choices such as linear affine vs. non-linear registrations, difference in registration cost-functions, and different kind of interpolations when applying the registration can all quantitatively impact. Does this make sense?
Thanks a lot for the quick and helpful answer, that does make sense!
Now I have a follow-up question for that: Since, in my case, most of these factors are predefined by templateflow’s ITK transform, my only choice to make here is the kind of interpolation.
Is there a “typical” or “generally recommended” kind of interpolation for registering probability maps? Or should I just go and see which one looks best at the end?
Personally, as long as you aren’t upsampling by more than 20% in the process, I think LanczosWindowedSinc is my go-to interpolation scheme (note that deterministic segmentations or label images should instead use nearest neighbor [or if ANTs, genericLabel]).
So in case anyone is reading this in the future and encounters the same problem, I’ll leave this additional comment:
I tried transforming the probability images using all the interpolation schemes of ants and compared the results to the standard MNI152NLin2009cAsym T1 image.
What seems to work best here is LanczosWindowedSinc, cosineWindowedSinc or linear.
In the end, it’s a very small difference, and you probably shouldn’t spend as much time on it as I did…
