Generating a wm mask out of multiple scalars

[36000]

No description provided.

[pep8speaks]

Hello @36000! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2020-08-10 15:40:14 UTC

[36000]

I think the documentation error here was just a connection problem.

[arokem]

I have a button that allows me to rerun builds. Maybe you see that one too?

[arokem]

Is this ready? The PR still says "WIP", so I figured I'd ask before merging.

[36000]

Nope! Right now I am using it on the prek data and I am finding that there is still overlap between the resulting wm mask and high values in MD. I marked this as WIP when I realized this was happening! I am investigating this bug and I will let you know when I think this works as intended.

[36000]

I think I have located the problem but I would like it if @arokem could make sure my understanding of this is correct.
We use this line on our wm_masks:
wm_mask = binary_dilation(wm_mask) > 0 (

pyAFQ/AFQ/api.py

Line 865 in 390ade6

wm_mask = binary_dilation(wm_mask) > 0

)
When we were getting wm masks from segmentation files / an arbitrary FA cutoff this made sure we were not being too strict. But now that we are trying to make specific masks we should not be doing this.
Should we do this for seg files but not wm masks based off of scalars? Or should it be based on user input?

P.S. Some extreme tunnel vision that I did not find this earlier...

[arokem]

Oh - good spot! I think that this was a precaution against overly conservative masks, but I think that's really not an issue anymore. Given everything we've seen in the last couple of weeks, I am starting to think that maybe there is never a reason to use binary dilation here. And as you've demonstrated, it can lead to some puzzling outcomes. Maybe we should remove it altogether?

[36000]

Because of this interpolation trickiness, I may just revert to using an FA mask for tracking to replicate mAFQ, but then mask out any nodes with high MD after profiling.