Hi,
I'm using FLIRT to register T1 structurals to avg152T1. If I use full
12 dof affine, then the results look beautiful. However, for
complicated reasons, I'd like to just use 6 dof rigid transforms, and
FLIRT seems much less successful at this (all same options except
dof; mostly defaults, e.g. correlation ratio, etc).
I suppose this makes sense, because the subjects don't match avg152
well in terms of overall brain scale (they are elderly/Alzheimer's) or
local things such as ventricular size. But is there anything I might
try to improve things?
I can't use a 12 dof as a starting point for a second 6dof run, can I?
Since I'll still end up with effectively 12 dof. I assume I can't just
take a 12 dof matrix, break it down into skew/scale/rot/trans and just
"reset" the skew and scale? I'm guessing this would give peculiar
results with e.g. the rotation that was derived to happen after some
scalings being off without them, but perhaps I'm wrong about this?
I guess since I'm doing T1-T1, I should perhaps be using normcorr,
rather than corratio? The contrast in my images didn't seem that close
to avg152T1, which was why I stuck with an inter-modal cost function,
but perhaps an intra-modal one would be better in the face of slight
imperfections due to the scaling problem? I gather (N)MI can have
quite a sharp basin in the cost function, perhaps corratio suffers
similarly, with slight mismatch being no better than yards away?
I open to any suggestions, however random they might seem!
Many thanks,
Ged.
|