Replace --anatomical-template and --output-resolution with --output-spaces

[tsalo]

This would bring QSIPrep more in line with other preps, and would maybe make it easier to support a res-native option for cases where resolution varies across acquisitions, sessions, etc.

@mattcieslak pointed out that most recon methods require isotropic voxels, so we might need something like res-nativemax and res-nativemin to use the largest or smallest voxel dimension, respectively.

For example, if voxels are 1.7 x 1.7 x 1.718, res-nativemax would produce outputs with 1.718 x 1.718 x 1.718 voxels and res-nativemin would produce outputs with 1.7 x 1.7 x 1.7 voxels.

[tsalo]

Satra proposed something like res-isoXmm for arbitrary isotropic resolutions in nipreps/fmriprep#2424 (comment).

[tsalo]

I've opened nipreps/niworkflows#889 to request this on niworkflows' side.

[tsalo]

We should require ACPC as an output space, and other output spaces won't use the resolution component.

So you could do something like:

--output-spaces ACPC:res-2:res-nativemin:res-nativemax MNI152NLin2009cAsym MNI152NLin6Asym

That would output preprocessed DWI in ACPC with three different resolutions, and would also generate transforms to MNI152NLin6Asym and MNI152NLin2009cAsym.

[tsalo]

I think this proposal can take over for #620. We could allow other output spaces than ACPC and raise a warning/error if someone wants something that doesn't make sense.

[tsalo]

A note on infant processing- MNIInfant without a cohort entity should resolve to the appropriate cohort based on the subject age. If the cohort entity is present, then we use that. Maybe we should have an 'auto' option as well? E.g.,

--output-spaces MNIInfant:cohort-1:cohort-auto

That would produce transforms to MNIInfant:cohort-1 and transforms to whatever cohort is most appropriate for the participant based on their age.

[mattcieslak]

One other thing I was thinking about - what if we removed options for spatial normalization from qsiprep entirely? The --output-spaces argument is misleading because the only space you'll get out is ACPC. It really just computes the registrations to each of those spaces, and those later get used by qsirecon.

It seems like it would make sense to have the normalization happen in qsirecon, where there will actually be scalars getting warped to that output space.

I had this thought during The Great Split, but decided against it because there are some practical problems with it. The first is, if I want to run separate qsirecon workflows on the same input data, I would end up computing the normalization multiple times, which would introduce error/variability across the qsirecon runs. The other issue is that for synb0-disco to work correctly we need a normalization to one of the MNIs.

[tsalo]

It would simplify the interface and make QSIRecon's data ingression easier. We can always allow QSIRecon to ingress derivatives from previous QSIRecon runs (see PennLINC/qsirecon#111), so the transforms wouldn't need to be recalculated.

I just worry about compatibility with other post-processing workflows. Minimal preprocessing should preferably produce self-contained outputs. It would be good to check with some of the other dMRI folks to make sure QSIPrep derivatives would still be useful without thoe transforms.

[mattcieslak]

As of now, those transforms are only used for the connectivity matrices. They used to be used for PyAFQ as well, but a recent PyAFQ update made the format incompatible with PyAFQ, so it just recalculates its own transform.

[mattcieslak]

correction: connectivity matrices and getting the ROIs for the ODF plots in the QSIRecon reports

[tsalo]

One reason to keep it in QSIPrep is that we can probably integrate sMRIPrep at some point. QSIPrep could do the native-to-ACPC transform and then pass the ACPC-space anats to sMRIPrep for full processing, right? sMRIPrep would produce the transforms to whatever output spaces we might want. Of course that could be done on QSIRecon's side, but I think it makes more sense to run Freesurfer in the preprocessing stage than the reconstruction stage.

This is kind of a dumb question, but how do you get tracts in the same space to look at patterns across subjects? Does that not use the ACPC-to-MNI transform?