ENH: Improve support of 4D in sdcflows.interfaces.bspline.ApplyCoeffsField

[oesteban]

This PR revises the implementation of the core unwarping interface, by enabling sdcflows.transforms.B0FieldTransform to project the coefficients representing the fieldmap into each volume of the input through the head-motion correction matrix, thereby moving the fieldmap with the head motion of that frame and hence applying time-step-wise correction.

I'm not confident this will really add substantial accuracy, but theoretically is the way to go. Given the sophistication (and the pervasive interpolation for each frame of the bspline field), this is not implemented elsewhere (except, perhaps, for applyTOPUP).

Resolves: #233.
Initiated-by: nipreps/fmriprep#2210.

[oesteban]

Okay, this is starting to look pretty good. I'm testing the new code within fMRIPrep and the feeling is optimistic.

It is very cool that B0FieldTransform is capable of producing hmc-ed and sdc-ed time-series, with a single interpolation step. Extension to sampling on surfaces and/or having further transforms to pre-concatenate does not seem like daunting future iterations at this point.

I might need to add some little changes here and there as fMRIPrep continues, but overall this is ready for your judgment.

[codecov-commenter]

Codecov Report

Merging #234 (8ca1e79) into master (1620310) will decrease coverage by 3.77%.
The diff coverage is 73.91%.

@@            Coverage Diff             @@
##           master     #234      +/-   ##
==========================================
- Coverage   93.66%   89.89%   -3.78%     
==========================================
  Files          24       25       +1     
  Lines        1673     1771      +98     
  Branches      191      205      +14     
==========================================
+ Hits         1567     1592      +25     
- Misses         81      150      +69     
- Partials       25       29       +4     

Flag	Coverage Δ
travis	85.31% <71.73%> (-3.63%)	⬇️
unittests	90.15% <73.91%> (-3.37%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
sdcflows/workflows/fit/pepolar.py	100.00% <ø> (ø)
sdcflows/transform.py	91.26% <70.83%> (-6.36%)	⬇️
sdcflows/interfaces/bspline.py	87.38% <71.18%> (-6.34%)	⬇️
sdcflows/workflows/apply/correction.py	100.00% <100.00%> (ø)
sdcflows/conftest.py	0.00% <0.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 1620310...8ca1e79. Read the comment docs.

[effigies]

This overall looks sensible to me. Having a bit of trouble keeping track of which way the HMC transforms are going. I kind of expected more affine inversions, but I'm assuming that you're keeping proper track...

[oesteban]

The following steps are:

i. Adding support to prepend additional transforms (e.g., epi-to-anatomical, anatomical-to-standard)
ii. Modulate intensity through the Jacobian of the SDC transform, prior to resampling
iii. An interface to project surfaces (requires (i))

As the three steps require the foundations set in this PR, I think the best path forward is merging this one. I'm aware I'm reducing coverage by ~4%, so I'll think about adding meaningful tests.