[Insight-users] Scaling in Registration

[Michael Hardisty]

I am attempting to use a centred affine transform to register two 3D 
volumes.  I believe that the registration should be a combination of 
shearing, scaling, rotation and translation.  Therefore the Affine 
transform seems like the ideal choice.  The problem I am having is that 
when I do the registration I never get proper scaling of the volumes.  
The Image I am registering is smaller than the target image and remains 
smaller after the registration has converged.  The other parts of the 
transformation look relatively good, the scaling is the only part of the 
registration that seems totally off.  I thought that the scaling might 
be suppressed by the particular metric that I am using 
(MeanSquaresImageToImageMetric), but I am not sure.  This leads me to 
several questions:

1)  How do the metrics evaluate when the images have different pixel 
densities?  Does it make sense to suggest that by stretching the moving 
image over more target pixels that this could cause a greater metric 
value because more pixels are included?  If so which metric is the most 
suited to deal with this problem?

2)  Are there any other parts of the registration process that may 
effect the scaling of the moving image?  What other strategies could I 
take to encourage more scaling?

Current Registration Components:
Metric:         MeanSquaresImageToImageMetric
Interpolator:  LinearInterpolateImageFunction
Optimizer:     RegularStepGradientDescentOptimizer
Transform:    CenteredAffineTransform
Image Types: uCT (pixel type = float)

Thanks

-- 
Michael Hardisty
M.A.Sc Student
University of Toronto
Orthopaedic Biomechanics Laboratory
Sunnybrook & Women's College Health Sciences Centre