[Insight-users] mutual information measure

Luis Ibanez luis.ibanez@kitware.com
Fri, 03 Jan 2003 12:44:33 -0500


Hi Jorn,

The loop of iterations on Registration methods is controled
by the optimizers. Metrics are used as the cost function to
be optimized.

It may be easier to get the value directly from the Optimizer,
just using its "GetValue()" method.

Note that only the value returned by the optimizer can be taken
as an indication of the evolution of the registration method.
Printing out the values computed by the metric can be misleading
because some optimizers compute derivatives by finite differences.
This implies that the metric "GetValue()" method is called several
times just to compute the derivative.

The "GetValue()" returned by the optimizer is the values that
is actually used to control convergence and eventually stop the
optimization process.

BTW,
The sensitivity to parameters is unfortunately quite common
on registration. There is a need for smarter methods to figure
out good combinations of parameters. You could imaging neural
networks or maybe fuzzy logic systems built on top of registration
methods for a next generation of software.


Please let us know if you have futher questions,


    Thanks


       Luis



-------------------------------

J. Van Dalen wrote:

> Dear itk-users,
> 
> I use itkMutualInformationImageToImageMetric to perform a registration. It
> all seems to work: I do not get error messages and by eye the two images
> seem to be on top of each other. However, the result depends very much on
> the value of the inputparameters (translateScale, numberOfIterations
> learningRate, standardDeviation, numberOfSamples). Now, I would like to
> tune these parameters to get an optimal result. Therefore, I would like to
> have some kind of a mutual information measure, that indicates how well
> the registration is performed. I already try to use GetValueAndDerivative
> in itkMutualInformationImageToImageMetric, but it doesn't give me any
> relevant information (I only observe fluctuations after each iteration).
> 
> Has anybody an idea? Is there something available in the itk routines
> that I overlooked?
> 
> Regards,
> Jorn.
> 
> 
> 
> _______________________________________________
> Insight-users mailing list
> Insight-users@public.kitware.com
> http://public.kitware.com/mailman/listinfo/insight-users
> 
>