[Insight-users] convergence of curvature registration

[Ruben Schilling]

Hi all,
I am experimenting with Thorsten Rohlfing's curvature registration filter in
the MultiResolutionPDE framework. Unfortunately there seems to happen
something wrong: I am estimating an affine pre-registration (with some other
algorithmic setup), calculate my own deformation field from that and set it
via the "-> SetInitialDeformationField()" method for the multi res curvature
registration as initial value. What happens then, is a straight ascend (!)
instead of a descend in the metric. This seems to be very unreasonable. I
validated of course the affine field with an image warper and difference
image (in comparison with the direct resampling using the affine transform),
so my field is ok. Since affine transformations lie within the kernel of the
curvature smoother I can from the theory not see, why it should impose
trouble to the curvature algorithm, when it is initialized with an affine
transformation. Did someone experience that too?

Regards
Ruben


PS
The above described setting converges without trouble, if I just resample
the template image once using the affine pre-registration and then do not
specify an initial field at all. Of course loosing one resampling means
loosing some precision in the final result.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://public.kitware.com/pipermail/insight-users/attachments/20071125/cb5901fe/attachment.html