[Insight-users] registration initialization problem - probably a stupid error

[Richard Beare]

Hi,
I'm experimenting with initializing a registration procedure that I'm
testing on the brain proton density images used in the examples. I'm
trying to use moment information to initialize a centered rigid 2d
transform, but I'm not using the initializer because 1) it isn't
available in python and 2) when I'm using my real application I will
want to do some segmentation and use those results to initialize the
registration.

Basically I'm doing pretty much what the moments based initializer
does anyway - I use the image moments calculator to get centre of
gravity and principal axes. The 2nd and 3rd parameters become the COG
of the moving image, the 4th and 5th become the difference between the
COGs of fixed and moving images (translation estimate). I thought the
first parameter (which I think is the rotation about the centre
indicated by parameters 2 and 3) would be best initialized by the
difference in angle of the principal axes of each image. This almost
works, but is a a couple of pixels off, which in the case of the
artificially shifted and rotated data, seems to be too much and
probably indicates I've made an error somewhere.

My procedure is to threshold the images, dilate a bit and compute the
moments of that image. The dilation doesn't touch the edge. If there
is translation but no rotation the translation estimate works
correctly.

I'm sure I've missed something simple - can anyone see what?

Thanks