Add --masktype option to control adaptive mask method

[tsalo]

Closes #312. I'd like to look into R-squared again at some point, but I think this is a solid improvement to the adaptive mask.

Changes proposed in this pull request:

• Add a new CLI argument, --masktype to both tedana and t2smap. This parameter accepts one or more values, like --gscontrol. The valid values are "dropout" (the current method), "decay" (flag any echoes that do not decrease in mean value), and "none" (set adaptive mask value to all echoes, for testing and the like).
  • If you use more than one method, the function will just take the elementwise minimum value from all of the method-specific adaptive masks.
• Implement a "decay" method for the adaptive mask, which calculates an array reflecting the first echo where mean signal does not decrease from the last echo.
• Remove getsum parameter from make_adaptive_mask. We don't use getsum=False outside of tests, so we really don't need it.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 89.75%. Comparing base (f084dc4) to head (a0d5f79).
Report is 35 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1057      +/-   ##
==========================================
+ Coverage   89.68%   89.75%   +0.06%     
==========================================
  Files          26       26              
  Lines        3482     3505      +23     
  Branches      615      620       +5     
==========================================
+ Hits         3123     3146      +23     
  Misses        211      211              
  Partials      148      148              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[handwerkerd]

Mostly looks good. I do want several people to compare the masks on real data to make sure nothing unexpected is happening. This change really shouldn't shrink the masks much, so, if it does, that's concerning.

[handwerkerd]

Might want to make a second warning message if last_decreasing_echo removes nontrivially more voxels than the other threshold. I'm viewing that as a sign that something is wrong with the data.

[handwerkerd]

I think we'll want to log the voxel counts with each different method. I wrote some code for that, but it was a bit ugly and I think it might be easier to do as a separate PR once all current adaptive mask PRs are merged.

[handwerkerd]

Thinking this over more: The code for these changes in simple, but this can non-trivially affect results in ways that are worth testing. One option to consider is that you tweak this PR to make it possible for an end-user to try the current adaptive mask, and several other options using t2smap (This could be hacky change to a couple of functions and the t2smap CLI/API, I could see either removing all of that or just the CLI/API option before merging.). It could also be tweaked to log the mask sizes & size differences with each option. Then it would be easy for several of us to run t2smap on a lot of data and compare the adaptive masks.

Depending on timing we could attempt to either have results to discuss by next Thursday's dev call or prod people to try this at the dev call. Thoughts?

[tsalo]

Based on our developer meeting a couple of weeks ago, I've added a --masktype parameter to control how the adaptive mask is generated. @handwerkerd could you take another look at some point?

EDIT: I've also made the old behavior (--masktype dropout) the default, so this should no longer change the results.

[handwerkerd]

I have a bunch of comments, but, overall, this is good. The only major thing is that I don't think it will be possible to select decay and dropout through the CLI.

Also, after starting this, I realized it would be slightly nicer to merge #1060 first since you'll pretty much have to re-write #1060 to fit with these changes. Changed my mind on this. It would be easier to merge this with the updated doc strings and parameters and then align the other masking PRs to fit within these updates.

[tsalo]

I think I've addressed everything, and I updated the t2smap integration test to use both dropout and decay to ensure that the parser could handle multiple values.

[handwerkerd]

I noticed a few more things, but this is getting close.

[handwerkerd]

I think we'll want to log the voxel counts with each different method. I wrote some code for that, but it was a bit ugly and I think it might be easier to do as a separate PR once all current adaptive mask PRs are merged.

[handwerkerd]

LGTM!

[eurunuela]

LGTM! Thank you @tsalo!