Dear FSLers,
I have been attempting to run sub-cortical segmentation using FIRST, using a (large) variety of methods on set of high-resolution, brain extracted 7T MP2RAGE images. Some examples:
run_first_all
first_flirt followed by run_first on each structure
flirt followed by run_first
The whole procedure appears to fail at the FLIRT stage, which misaligns the brains very badly (often inverting them, or shrinking them to the size of a walnut). This appears to be due to FLIRT being less robust when the reference image is lower resolution/contrast than the input image, but this does not seem to be the whole story.
Namely, the best results I achieved have been using the last method, where I back-normalise the (BETted) T1 template to the (BETted) individual brain, and then re-flirt using the sub-cortical weighting mask. This gives a significant, but not sufficient improvement: it fails in 20-30% of brains, rather than 40% when simply using run_first_all.
One suggested option of using robustfov from fsl_anat to crop the individual brains unfortunately does not help at all (if anything it makes things worse).