[FIX] Change to Linear for upsampling

[mattcieslak]

Addresses #819

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 0% with 2 lines in your changes missing coverage. Please review.

Project coverage is 28.12%. Comparing base (aa43d3b) to head (befeefb).

Files with missing lines	Patch %	Lines
qsiprep/interfaces/images.py	0.00%	2 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #820   +/-   ##
=======================================
  Coverage   28.12%   28.12%           
=======================================
  Files          69       69           
  Lines       10079    10079           
  Branches     1302     1302           
=======================================
  Hits         2835     2835           
  Misses       7146     7146           
  Partials       98       98           

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

[tsalo]

Why do you use Lanczos windowed sinc for downsampling and upsampling to >90% of the original voxel size. Would it make more sense to just use linear no matter what?

[mattcieslak]

Why do you use Lanczos windowed sinc for downsampling and upsampling to >90% of the original voxel size. Would it make more sense to just use linear no matter what?

Lanczos Sinc is very accurate if certain conditions are met. As long as there aren't sharp cutoffs between high and low intensity and the image resolution in the output isn't higher than the input data it is a great default (and is the default in fmriprep). A added the 90% as some wiggle room for weird voxel sizes.

Linear intepolation ends up smoothing the data, which is not something we want to do unintentionally. I think it makes sense for upsampling, since it won't introduce the ringing effects like a Sinc interpolator will.

I didn't think BSpline would add ringing (particularly the relatively simple BSpline[3] we were using), but in some cases it looks like can. Although, this may only happen if there is significant upsampling. I'm not sure what the native resolution was of the data that showed the ringing artifacts.

[cookpa]

LGTM