[Insight-users] about MIThinPlateSplineRegistrator

[Luis Ibanez]

Hi zhang,

That's a good idea, this is a plausible
approach for deformable registration.

As you correctly notice, the missing piece
is the SetParameter()/GetParameters() method
in the KernelTransform.

I'll add that and complete the example in:

Examples/MutualInformationEuler2DRegistration

for:    MIThinPlateSplineRegistrator


The choice of what the parameters are, is in
some way arbitrary. We could go easily by
taking the coordinates of the landmarks as
the parameters of the transformation.

For the case that you mention: 10*11 landmarks
in 2D, that will result in a array of:

  10 * 11 * 2 = 220 parameters.

which will be just the sequence of {x,y} pairs of
all the landmarks.

These parameters will be passed to the optimizer.

We could interpret the optimization as an iterative
process in which each landmarks is moved slightly
around its current position and the resulting
transformation is computed in order to evaluate if
the new mapping between the two images results in a
better matching.

You don't actually need access to the W matrix
since the matrices are computed directly from
the landmarks. Using the W matrix coefficients
as parameters could be difficult since it is
not clear that a variation of a W matrix is
neccessarily a valid matrix for a new set of
landmarks.

Moving the landmarks and recomputing the W matrix
for each new combination seems to be a more
intuitive way to go.
(and an easier one to implement :-)


I'll add the Set/Get parameters to the Kernel
transfom and get back to you soon.


Thanks

   Luis


========================================

zhangzhijun wrote:

> Dear Luis:
>          I want to put the TPS, EBS and VS transformation into the
> multi-modal registration frame work
> it is implemented following the 3 steps--- transform, similarity metric and
> optimization, now since the
> example of MIThinplatesplineRegistrator is not complete, I think I need do
> the following work:
>  1. Add an Initialization of the parameters for the TPS, in my case I want
> to register two
> 2D image of 181*217 dimensions, I choose totally 10*11 landmarks for
> calculating of the TPS parameters,
> After I calculate the W matrix, the itkKernelTransformation reorgnized the
> Wmatrix into three part but the
> 3 matrix members are all protected member, I find I can't have access to it.
> I want to set the 226 parameters
> and then use optimization method to do the optimization.
>  2  also I don't know how the deviation in the itk is implemented and what's
> the order for these 226 parameters.
> because I need do an steepest descent optimization, so the order of every
> parameter should follow some convention,
> I don't know what is it.
> 
> regards,
> zhang zhijun
> 
> _______________________________________________
> Insight-users mailing list
> Insight-users@public.kitware.com
> http://public.kitware.com/mailman/listinfo/insight-users
> 
>