[FIX] INU correction before merging several T1w

[oesteban]

Otherwise, we get funky composite bias fields that N4 can't recover.
This will also make coregistration between T1s robuster.

Fixes #924

[effigies]

This looks reasonable, though I think we can simplify further in the N = 1 case.

My main concern will be whether doubling up of N4 correction (because it's also performed in the skull-stripping step) might cause artifacts.

Possibly a simpler (conceptually, if not in practice) solution would be to factor N4 out of skull-stripping, apply before merging, and eliminate the need to branch on number of T1w images at all.

[effigies]

t1_reorient is not necessary if t1_merge is not used. Conformation orients to RAS, but for reasons I don't understand, merging sometimes changes the orientation. Then you can get rid of concat_affines and just use the conformation affine.

[oesteban]

I've been working at times in a pure-nipype implementation of antsBrainExtraction.sh. However, I haven't got any decent prototype yet (lack of time).

I'm not particularly worried about running N4 twice though, as long as it gets us sensible brain masks. But I agree, it'd be better not doubling up.

[effigies]

Okay, so if we're fairly confident that the quality of the brain mask would be unaffected, but don't want to risk any artifacts of doubling up N4, we could use the following sequence in all cases:

• Conform
• N4-correct
• Merge
• Skull-strip
• Apply brain mask to merged image

This would mean we stop using the BrainExtractionBrain and N4Corrected0 outputs of brain extraction directly. We could slide the extraction into the skullstrip_ants_wf to keep the interfaces consistent.

WDYT?

[oesteban]

Yes, though I would give it a try to what we have right now.

My real, deep interest in deconstructing antsBrainExtraction.sh is to be able to connect stuff in the middle, without modifying the nipype interface, and to allow for tweaks. Great things we could have:

• Find a mapping from OASIS (or NKI) to MNI and initialize spatial normalization with the precomputed transform of brain masking (remove 20 min from runtime).
• Run the brain extraction on all T1s separately:
  1. Run an antsTemplate -like workflow to replace the fs.RobustTemplate.
  2. Initialize second and subsequent transforms with the first OASIS-T1w transform followed by the affine found in i).
  3. Final brainmask is the consensus of all masks.

But I can't see this happening anytime soon.

[effigies]

Alright. So we can keep this if branch for now, and maybe create an issue to move to avoid having double N4 correction in skull-stripping outputs. It shouldn't be a huge task, but it will still be some work.

[oesteban]

I'd go for that. I'm simplifying the workflow following your suggestions.

[effigies]

Looks reasonable. Some small comments.

[effigies]

Should go in a "Next release" section.

[effigies]

I think this file will need to be added to setup.py?

[oesteban]

yes

[effigies]

I think you can revert lta_to_fsl and concat_affines entirely, cleaning up the diff a little.

In any event, this no longer needs the 1 + min(...) construct.