Hi Matt,
> So I take it that setting -ds to a lower number will result in more control points
Yep, essentially I think df = const / ds^3 for a particular image
size. Similarly registration time probably goes as ds^(-3), which can
make fine cp spacings quite painful...
> I wonder does anyone know what the optimum control spacing for
> registering high-resolution structural scans to a template
Deciding how good registrations are is not trivial (have a look for
the "Zen and the art of ..." TMI paper by Bill Crum), and things like
optimum cp spacing will depend heavily on the particular anatomy of
the brains, the particular imaging sequences, the quality of the
initial affine alignment, etc. etc.
I think most stuff I've seen has plumped for 20mm as a good guess for
inter-subject registration. For intra-subject I've seen anything from
20 down to 1.something reported, though I've heard (second hand) that
DR recommends 2.5 times the voxel size as a minimum spacing.
Note that you can use the -subdivide option with -dofin to halve the
resolution of a previously estimated transformation. You'd have to
check with Daniel to get the exact syntax (I use a slightly different
version of nreg to that available in FSL, and I don't even have access
to that at the moment, so can't be very helpful, sorry). This lets you
follow a multi-resolution approach, e.g. stepping the cp spacings from
20 to 10 to 5 to 2.5, which I think is quite common within DR's group.
There are then some questions about what blurring to use at the
different levels (take a look at nonlinear.par etc.) for the best
performance. I'm afraid I can't give you a good answer for that...
Subdivision gives much better results than simply using the finest
spacing to start with; doing that would almost certainly lead to
finding local minima in the cost function, and probably to folding of
the deformation field (negative Jacobian), which would be fairly
disasterous for VBM (I'm not sure whether TBSS involves modulation by
the Jacobian), and would prevent inversion of the transformation
(though I don't think FSL's IRTK provides a tool for this yet anyway).
Best of luck,
Ged.
|