[Insight-users] Re: FEMRegistrationFilter

[Luis Ibanez]

Hi Andy,

When you use the FEMRegistrationFilter for performing
registration you should think of your images as a block
of rubber on which you are applying forces.

If the rubbers doesn't deform much, it means that your
forces are not large enough for the elasticity coefficient
of the rubber.

You have then two options


       A) Change the rubber physical properties

       B) Increase your forces.



Since the forces are computed from the derivatives of
the metric, you don't have much of an option in (B).

Therefore your best shoot is to play with the values
of Elasticity and Density that define the physical
properties of the material.


Please use the methods:

               SetElasticity();
               SetRho()

The larger the Elasticity, the lower the deformation.
This is just Hooe's law

                F   =  E   *   Dx

Dx = deformation
E  = Elasticity
F  = Force


The following is the advice in the Doxygen documention
of the FEMRegistration filter:


http://www.itk.org/Insight/Doxygen/html/classitk_1_1fem_1_1FEMRegistrationFilter.html


  "Optimal parameter settings require experimentation. The following
  approach tends to work well : Choose the relative size of density to
  elasticity (e.g. Rho / E ~= 1.) such that the image deforms locally and
  slowly. This also affects the stability of the solution. Choose the
  time step to control the size of the deformation at each step.
  Choose enough iterations to allow the solution to converge (this may
  be automated)."


Regards,


    Luis


---------------
Andy Eow wrote:

> Hi Luis,
> 
> I'm interesting in performing a baseline deformable registration of a
> 256x256 image with its chiral using a MI metric. Currently, I'm
> experimenting with using the FEMRegistrationFilter and configuring the
> filter via code instead of a configuration file. In any case, I'm not
> exactly sure what parameters should be changed and by how much for the MI
> case.
> 
> registration->ChooseMetric(3);
> registration->SetDescentDirectionMaximize();
> registration->SetMaximumIterations(500,0);
> 
> These are the only differences from the MS set of parameters that I've
> changed so far. In this case, the filter returns virtually no deformation
> after registration even though other filters like DemonsRegistration showed
> significant deformation in certain regions. Code works relatively well when
> I use the Means Square metric option instead as shown in the sample
> configuration file. Hence, I'm guessing that there's still some parameters
> that should be changed accordingly. Please advise.
> 
> Thank you very much.
> 
> Cheers,
> Andy
>