pyAFQ pipeline: Use previous analysis derivatives as inputs for masks: seed and brain.

[arokem]

Changes proposed in this pull request

This changes the defaults for seed mask and brain mask for the pyAFQ pipeline, using pyAFQ's mask definition API.

Might resolve #563

[36000]

The same change should be made to this file: https://github.com/PennLINC/qsiprep/blob/master/qsiprep/data/pipelines/mrtrix_multishell_msmt_pyafq_tractometry.json

[arokem]

Hey @36000 : should this use an ImageFile or a MaskFile object? Would this work as written here? That is, do these names (ImageFile, MaskFile) exist in the namespace in which this config is run?

[arokem]

Also, can anyone tell me how mrtrix pipelines choose where to place seeds? In particular, the msmt_noACT pipeline?

[arokem]

@mattcieslak : what should we put here for wm_mask?

[arokem]

Updated (with help from @36000, thank you!) to do this the right way. The config is now explicit about using the WM_prob output to avoid confusion (it turns out that I was wrong in #563, because the default settings use FA with 0.2 as a threshold, but that logic is a bit hidden in pyAFQ, which we will also fix separately...), and a probability of 0.5 as the threshold.

[mattcieslak]

I think some of the recent changes in the anatomical workflow are causing some of the data AFQ previously used to not be present anymore. I'm definitely open to suggestions on a best default behavior here, but I really like the idea of using a low FA threshold instead of an anatomically defined WM mask. This is because the anatomical masks are unreliable because they can come from a couple sources.

• If 0.17 or earlier is used there may be a 3-tissue dseg and a probseg from FAST. These are almost always bad.
• QSIPrep After 0.17 will have a 3-tissue dseg from SynthSeg. These are usually excellent, but it's still a new method and I have seen some bad failures with it.
• For recon-only you may have a great segmentation from non-synth FreeSurfer available. This is also just a deterministic segmentation, though.
• If the user ran a --dwi-only you have a mask based only on a mostly reliable b=0 masking method.

Also if the user doesn't have fieldmaps available the t1w or t2w-based tissue masks won't align well with the processed dwi data.

For the mrtrix hsvs we require a freesurfer segmentation, which guarantees good results as long as susceptibility distortion was run. We could impose a requirement that a good anatomical segmentation from freesurfer is available if the user wants to use it for AFQ. It seems simpler and (if no fieldmap is available) more effective to use a FA mask.

What do you think?

[tsalo]

Is this something we should reproduce in QSIRecon?

[arokem]

Yes, that would be great. We routinely do this kind of thing when we run pyAFQ on top of qsiprep derivatives, and it's beneficial. The default for brain-masking in pyAFQ, in particular, is quite poor as it relies on the (not very good) median Otsu algorithm for brain extraction.

[tsalo]

I've opened PennLINC/qsirecon#170 to track this in QSIRecon. I'll close this PR since there's no way to merge it in QSIPrep but I have a link to it from the QSIRecon issue so someone can copy over the changes to their own branch at some point.