ENH: Add a --skull-strip-t1w argument to skip brain extraction

[oesteban]

Supersedes #1829.

@adelavega can you check locally this works for you?

[mgxd]

lgtm

[effigies]

Why is the default not "auto"?

[adelavega]

I'd think auto should be the default as well, otherwise if a user is changing that option its likely they know they need to skip brain extraction. That said, I can understand if we don't think the algorithm is reliable enough (yet?)

Will try this out soon

[effigies]

Some alternative wordings, if we do switch to auto as default (which I think we should).

[effigies]

Suggested change

	skull_strip_t1w = "force"
	skull_strip_t1w = "auto"

[effigies]

Suggested change

	'--skull-strip-t1w', action='store', choices=('auto', 'skip', 'force'), default='force',
	'--skull-strip-t1w', action='store', choices=('auto', 'skip', 'force'), default='auto',

[oesteban]

I'd rather keep the 'force' mode as default, since it is currently the "official" operation mode. This flag was suggested to cover a particular use case (when the user is not aware of having T1w images already masked), and crashing first seems like a conservative but effective approach.

Finally, this feature should be superseded by evolutions of nipreps/smriprep#107 (and #2036) where one can feed fMRIPrep with subsets of precalculated anatomical derivatives.

Is that convincing?

[effigies]

I don't see the point of having auto at all, then. The idea was to have something that didn't require using a CLI flag most of the time.

[oesteban]

Removing 'auto' would effectively be the same as #1829 (i.e., a boolean switch). However, when writing this I thought it was worth having 'auto' for when your dataset has both full and brain-extracted T1w, and you don't know which participants while you trust the heuristic. I can add a warning saying "use with caution" to emphasize this point.

Without a benchmark that says the heuristic works with very high sensitivity and specificity, I feel resistant to making it the default.

[oesteban]

For instance @nicholst et al., are processing ds001 and have found fMRIPrep gives them very bad results for sub-04.

Looking at the reports, they discovered that the T1w is skull-stripped.

I think having an implicit option would have been potentially dangerous to their type of analysis because it would have been really easy to miss that fMRIPrep actually ran different workflows for that subject w.r.t. the others (and this is assuming 'auto' worked perfectly on all participants).

Setting the default to 'force' (i.e., the original fMRIPrep behavior before this patch) is more explicit.

[nicholst]

Setting force as default makes sense for continuity of usage. If people get some experience with auto and find it does a good job catching extracted brains, then later it could be switched to auto.

[adelavega]

Works for me. FWIW I just tested in on ds001110 with auto and it worked well (detected skull stripped T1 and skipped the workflow).

[oesteban]

Thanks for the fast check and your patience, @adelavega. This took a lot longer than we would've liked (for the most part, because of me) - many thanks.