experiment comparisons

[jordandekraker]

As discussed in the grp-brainscores meeting, @rcruces and I had these comparisons to suggest (Raul feel free to add if i missed anything):

• LamaReg (labelsonly)
• LamaReg (robust - likely the new default)
• ANTs default
• ANTs robust (add highres iterations only, even though lowres initialization will still fail)
• ANTs steelman (no lowres iters, robust highres iters) (also note this may actually be WORSE for some datasets that are low res)
• fMRIprep defaults (if different from above)

You may still get additional Reviewer requests, but I think this is solid coverage as a start point.

Also when it comes to test datasets, focus on generalizability rather than within-dataset precision. Maybe only keep 20-50-ish subjects per dataset (that should be enough for a robust estimate of registration success), but try to cover a few broad cases:

• 7T (PNI)
• 3T research-quality (MICs)
• clinical quality (low res, maybe talk to @ella-sah about a good dataset to use)
• possibly some HCP1200 (since its so widely used)
• possibly a disease case, like ADNI

[Ian-Goodall-Halliwell]

Also MRTriX and FSL to test registration

[Ian-Goodall-Halliwell]

Actually, MRTriX is only really designed to do intra-DWI registration

[Ian-Goodall-Halliwell]

Also, fMRIprep uses either FSL's epi_reg or Freesurfers's bbregistration, so I should be covered using just FSL's epi_reg for now I think, in addition to the ants robust

[rcruces]

It looks like with those will cover most of the traditional registration methods in neuroimaging. I would include easyreg on thewhich uses the same first step as us (synthseg) and a CNN for the registration.

see:
https://surfer.nmr.mgh.harvard.edu/fswiki/EasyReg