[Insight-users] What optimizer to use for an affine registration and why?

[Demian Wassermann]

Hi to all,

I'm trying to implement an affine registration app as a piece of a pipeline. So far in a lot of cases the result is good but the affine matrices I get are crazy.

I'm testing simply with a 3D diamond centered. Using the normalized correlation metric. And I apply to the diamond either a translation, a scaling or a rotation. Everything is centered in the image (and the affine transform too).

By crazy I mean that, for an anisotropic scaling along the axes (meaning I apply a transform of a diagonal matrix with (0.8, 1, 1.2) in the diagonal) and then try to recover this through registration, the resulting matrix is a much more complex linear transform.

I think this is due to the choice optimizer and local minima. Any recommendations on this?

Thanks!
Demian

--
Demian Wassermann, PhD
demian at bwh.harvard.edu
LMI / PNL / SPL Labs
Harvard Medical School
Brigham and Women's Hospital
1249 Boylston, Boston, MA, USA