[Insight-users] jumps in MSE for multiresolution registration

[Ramón Casero Cañas]

Hi,

I have implemented a basic registration algorithm following the ITK 
Guide examples, I get "jumps" in the MSE metric when changing scales, 
and I'm not sure whether this behaviour is to be expected.

Before getting to code details, I would like to present an overview of 
the problem.

I am using the:

  * MeanSquaresImageToImageMetric
  * LinearInterpolateImageFunction
  * MultiResolutionImageRegistrationMethod
  * RegularStepGradientDescentOptimizer
  * TranslationTransform

I am registering two MRI hearts. I have attached an EPS graph with the 
error metric over the number of iterations.

The multiresolution scheme uses 3 leves, with the default schedulers. In 
the figure, one can clearly observe jumps in the MSE metric when the 
scale is changed.

In this case, at iteration 60 we are in Level 0, and the MSE = 2.19e+07. 
At iteration 61 we are in Level 1, and the MSE = 2.65e+07.

But given that the MSE normalizes by the number of voxels, shouldn't the 
transition between scales be smooth? Is this likely to be just a bug in 
my code?

(Fig. 8.41 in the ITK Guide shows metric values over iterations in a 
multiresolution scheme using the AffineTransform, but to me it is not 
clear where the level transitions are, and whether these "jumps" happen 
there too.)

Best regards,

Ramon.

-- 
Ramón Casero Cañas, DPhil

Computational Biology, Computing Laboratory
University of Oxford
Wolfson Building, Parks Rd
Oxford OX1 3QD

tlf     +44 (0) 1865 610807
web     http://web.comlab.ox.ac.uk/people/Ramon.CaseroCanas
photos  http://www.flickr.com/photos/rcasero/
-------------- next part --------------
A non-text attachment was scrubbed...
Name: multires-mse.eps
Type: image/x-eps
Size: 9473 bytes
Desc: not available
URL: <http://www.itk.org/pipermail/insight-users/attachments/20090715/f4b996b5/attachment-0001.bin>