Apply transforms to raw data when MNI6 derivatives aren't available

[tsalo]

Closes #57. I tested this locally against running with MNI6 derivatives, and the results are different. I am a little worried that I didn't implement the transforms properly or something.

To do:

• Ensure boilerplate is accurate
• Load STC slice ref from preprocessed BOLD data (need to find preproc BOLD data in any space). Moved to Infer the slice timing reference from fMRIPrep derivatives #60.

Changes proposed in this pull request

• Add ignore to the Config object so users can actually ignore stuff.
• Ignore fieldmaps by default, since I don't know how to grab them yet. Will revert this later.
• Add mem_gb to SUSAN and MELODIC nodes (closes Add mem_gb to SUSAN and MELODIC nodes to avoid memory errors #11).
• Allow processing of raw BOLD data when MNI6 derivatives aren't available.
  • We still need an anat-to-MNI6 transform.
  • This runs STC, then init_bold_volumetric_resample_wf, using raw BOLD and derivatives I think should be available at least in full mode with --cifti-output.

[tsalo]

Need to figure out how to grab the field map

[codecov]

Codecov Report

Attention: Patch coverage is 15.29412% with 72 lines in your changes missing coverage. Please review.

Project coverage is 31.08%. Comparing base (d3f8b26) to head (0754dbb).

Files with missing lines	Patch %	Lines
src/fmripost_aroma/workflows/base.py	0.00%	45 Missing ⚠️
src/fmripost_aroma/workflows/outputs.py	33.33%	18 Missing ⚠️
src/fmripost_aroma/interfaces/nilearn.py	0.00%	9 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #58      +/-   ##
==========================================
- Coverage   31.40%   31.08%   -0.32%     
==========================================
  Files          26       26              
  Lines        2277     2329      +52     
  Branches      360      362       +2     
==========================================
+ Hits          715      724       +9     
- Misses       1541     1584      +43     
  Partials       21       21              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[tsalo]

With MNI6

Without MNI6 (i.e., warping raw)

That first component really throws me.

[tsalo]

I realized I wasn't passing along the STCed BOLD data to the MNI6 workflow, so I fixed that, but it didn't bring the results in line with the MNI6-derivatives outputs.

[tsalo]

I finally took a look at the warped data from my PR and the data are in the wrong space. The MNI-space BOLD from fMRIPrep is in red and overlaps well with the brain mask. The warped BOLD from this PR is in blue and does not overlap with the brain mask.

[tsalo]

The transforms being used for the resample (ResampleSeries) step of bold_MNI6_wf (init_bold_volumetric_resample_wf) seem right to me:

1. sub-01/ses-01/func/sub-01_ses-01_task-rest_from-orig_to-boldref_mode-image_desc-hmc_xfm.txt
2. sub-01/ses-01/func/sub-01_ses-01_task-rest_from-boldref_to-T1w_mode-image_desc-coreg_xfm.txt
3. sub-01/ses-01/anat/sub-01_ses-01_from-T1w_to-MNI152NLin6Asym_mode-image_xfm.h5

Comparing to what's in the fMRIPrep working directory (they're the same!):

1. sub-01/ses-01/func/sub-01_ses-01_task-rest_from-orig_to-boldref_mode-image_desc-hmc_xfm.txt
2. sub-01/ses-01/func/sub-01_ses-01_task-rest_from-boldref_to-T1w_mode-image_desc-coreg_xfm.txt
3. sub-01/ses-01/anat/sub-01_ses-01_from-T1w_to-MNI152NLin6Asym_mode-image_xfm.h5

[tsalo]

Okay, I think it's an issue in the Docker image. I ran ResampleSeries locally and the results look good for boldref-to-T1w-to-MNI6 and orig-to-boldref-to-T1w-to-MNI6. I then tried it on the same files, with the same commands, in the Docker image, and, while boldref-to-T1w-to-MNI6 looks good, orig-to-boldref-to-T1w-to-MNI6 does not. So basically the HMC part is not operating correctly within the Docker image.

I assume the problem is buried in one of the Python dependencies in my Docker image.

[tsalo]

@effigies I bumped my local nitransforms from 23.0.1 to 24.0.0 and the orig-to-boldref-to-T1w-to-MNI6 result started looking bad. Is there some chance an issue with using fMRIPrep's HMC transforms was introduced in nitransforms 24.0.0?

[effigies]

Can you decompose that? I don't think anything should have changed in HMC, but we did fix the parsing of the ITK warps, which might have broken our workaround in nipreps...

[tsalo]

I can create a minimal reproducible example, probably.

Here's a more linear summary:

1. I ran fMRIPrep 24.0.0 on a single BOLD run with 100 volumes. I included func and MNI152NLin6Asym:res-2 as output spaces.
2. I pulled the fMRIPost-AROMA Docker image, which has nitransforms 24.0.0 installed.
3. I created a Python 3.10 environment on my laptop with nitransforms 23.0.1.
4. I applied the coregistration and normalization transforms to the func-space derivatives from fMRIPrep with ResampleSeries, both in my Docker container and in my local environment. Both produce reasonable MNI152NLin6Asym-space outputs.
5. I also applied the motion correction, coregistration, and normalization transforms to the raw BOLD file with ResampleSeries in both the Docker and environment. The Docker container (nitransforms 24.0.0) produced an output that looks like the blue brain in the above figure. The local environment (nitransforms 23.0.1) produced an output that looks good.
6. Then I updated my local environment's version of nitransforms to 24.0.0 and re-ran my ResampleSeries command. The output now looks like the bad Docker-produced one.

[effigies]

Okay. Damn. We probably need to roll back nitransforms in the fmriprep 24.1.0 branch.

[effigies]

Actually, would you mind re-testing with fmriprep 24.1.0rc0, to verify that the problem exhibits there as well?

[tsalo]

I can run the subject through the full pipeline later today, but in case it's informative, I pulled nipreps/fmriprep:24.0.0 and nipreps/fmriprep:24.1.0rc0 and ran just the ResampleSeries command with those. The results look good for 24.0.0, but not for 24.1.0rc0.

[tsalo]

I've got fMRIPrep running with 24.1.0rc0 and I'll update this thread with my results once it finishes.

The good news is that I pinned nitransforms to 23.0.1 and rebuilt the fMRIPost-AROMA container, and the results look solid to me, so I'm going to merge this PR.

[tsalo]

We're having server issues right now so it might be a few days before I can resubmit my fMRIPrep 24.1.0rc0 job.