Identify the last good echo in adaptive mask instead of sum of good echoes

[tsalo]

Closes #679. I built this off of #1060 and #1057, so it shares many changes.

This will only ever increase the adaptive mask value, as the last good echo will always be equal to, or higher than, the total number of good echoes.

Note that this does not identify the last contiguous good echo (i.e., the last good echo before any bad echoes). Therefore, if you have something like [good, bad, good, bad], the result with be 3, not 1.

Changes proposed in this pull request:

• Identify the last good echo for the adaptive mask, instead of the total number of good echoes.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 89.80%. Comparing base (461bc38) to head (2d3712a).
Report is 25 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1061   +/-   ##
=======================================
  Coverage   89.79%   89.80%           
=======================================
  Files          26       26           
  Lines        3537     3540    +3     
  Branches      620      621    +1     
=======================================
+ Hits         3176     3179    +3     
  Misses        212      212           
  Partials      149      149           

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

[eurunuela]

Changes look good to me.

[handwerkerd]

I was trying to get a better handle with what was happening within this PR. In make_adaptive_mask I added older_mask =(np.abs(echo_means) > lthrs).sum(axis=-1) right after dropout_adaptive_mask was calculated and compared results in my debugger. Calls like ((older_mask==2) * (dropout_adaptive_mask==2)).sum() show voxels that used to be 1 in the adaptive mask and are now 2.
Voxels that were 0 are still 0. (This interacts with other masking steps, but seems matched at this point in the code)
Of the 3590 voxels that were 1, 1061 are now 2, and 913 are now 3.
Of the voxels 5276 voxels that were 2, 1900 are now 3.
As expected, none of the voxels that had a higher values are now lower.

That means this change will substantively expand the number of voxels used in ICA and will balance out some of the drop caused by raising the threshold from masking. The one thing that concerns me is it turns how this test data had 913 voxels where the first and second echos were both below threshold, but the third was above threshold. That doesn't seem great, but the proposed enhancement would give this voxels a 0 in the adaptive mask, which would mean they'd be even be excluded from the optimally combined data. There's nothing to change here, but this is a discussion that will be relevant if a future enhancement is considered.

[tsalo]

I've opened #1083 about this. Can you follow up in that issue with your ratio idea?

[handwerkerd]

I suggested a code simplification, and a more detailed comment. There's one more comment that's more relevant for future discussion. Besides that, this should be ready to merge.