Using affine matrix

[dmcallindon]

I'm evaluating changing our lab's neuroimaging pipelines to use deepmriprep. I'm very impressed with it so far - installation was straightforward and results look amazing. I'm using command line.

One thing we often do is to transform a custom ROI from template space into subject space. I tried to use torchreg (with a new "prepare" function to set up the affine transformation matrix with the inits without doing a new registration, then the torchreg transform function) to transform between aal3 and aal3_affine (both ways), but neither quite worked. I checked that my affine matrix after initializing (using the individual rotation, translation, etc results) matches the deepmriprep affine matrix and I'm transforming using the same torchreg, so why is it not working? Which way is the affine matrix (i.e., template ->subject or reverse?) and which template are you using (e.g. FSL MNI152_1mm_brain?)

Thanks,
Dave.

[waigwamacha]

Hi @codingfisch, thanks for getting deepmriprep together, it's very impressive and we are using it at or lab as well. I am also curious to get an answer to @dmcallindon's question. Would be really nice to have the functionality to move from MNI space to subject space, especially for projects where precise localization is critical. Please let us know.

Thanks!

[codingfisch]

Hi @dmcallindon and @waigwamacha,
thank you for the nice feedback!
The atlas registration, and especially the transformation into subject space, is the most difficult to understand part of the pipeline defined in the run_atlas_register-method and the AtlasRegistration-class it calls. You can copy and paste these ~50 lines of source code into an LLM and let them explain to you but in short each original atlas nifti is...

1. padded such that it is affine-aligned to template space (MNI152NLin2009cAsym from CAT12)
2. warped using warp_yx to arrive at atlas_affine
   • warp_xy deforms from subject to template (for already affine-aligned niftis), so warp_yx does the inverse (template to subject=what we need here)
3. atlas_affine is moved with the inv_affine (template to subject) to arrive at the atlas in subject space

The best way to do the same with a custom atlas/ROI is to

1. clone this repo
2. add your atlas in deepmriprep/data/templates (also create the respective CSV if you want the atlas_volumes CSV output)
3. add the filename of your atlas in the ATLASES-dict of deepmriprep/atlas.py
4. run it eg via run_preprocess (see README)
5. if you want to have it as your new permanent deepmriprep run pip install . in the folder that contains the pyproject.toml

This way you have it nicely integrated and even in the GUI your custom ROI would show up as an atlas.

If your custom ROI can be expressed as one or a couple of ROIs in one of the already supported ATLASES you can just pick/merge these ROIs in the deepmriprep outputs as a quicker workaround.

I hope this helps you to fix the issue!

[dpmcallindon]

Hi @codingfisch ,

It's an elegant solution. Some experience that may help others or help refine the code:
I was getting error in line 52 of atlas.py in pad_from_to: ValueError: could not broadcast input array from shape (181,217,187) into shape (169,205,169). Seems to be the first step you mentioned - padding - is not working all the time. That was with an atlas made using the 1mm version of MNI152NLin2009cAsym, downloaded from source.
So I tried making my atlas using the WARP_TEMPLATE image Template_4_GS and taking, in AFNI terms, the first sub-brick of the template. Then it worked, although it seems making the csv file is not optional, unless it is only because I ran with the -ns flag to get error messages?

That's enough to get me on my way!