[Insight-users] Re: Using Paraview to understand deformable registration results

[Luis Ibanez]

Hi Barbara,

Thanks for your detailed email.


1) The deformable-registration experiment that
    you are running with the spheres is a bit
    extreme. Your current spheres do not have
    any overlap. It is therefore difficult to
    initialize the deformation field in such a
    way that it will capture their relative
    displacement.

    I wouldn't expect this registration to work
    if you don't have at least a 1 radius overlap
    between the spheres, and if you want to be
    more realistic, you probably should have 3/4
    diameter overlap.

    Deformable registration methods *are not*
    designed for accounting for large displacements.
    In practice you want to solve first any Rigid
    (or even Affine) misregistration between the
    two images *before* you pass them to a Deformable
    registration algorithm.


2) I would suggest you to start with an *easy* case.
    Like setting the gray sphere (the moving image)
    just 1/10th of radius off from the position of
    the red sphere (the fixed image).

    Fine tune the parameters of the deformable registration
    algorithm until you get reasonable results for this
    known translation.  Only then, you can dare to push the
    spheres farther away.


    WARNING : Note that deformable registration algorithms
    can give you answers that you don't want to hear.
    That means that in practice you may not necessarily get
    a nice horizontal displacement field because that's *not*
    the *only* way of morphing an image into the other.

    You could also get at answer such as

        Collapse half of the moving sphere into a point,
        and expand the remaining half of the sphere into
        a full sphere.   The displacement field for this
        solution will look like if you were squeezing the
        sphere to make it pass through a hole.

     Here is where the assumptions of your physical model
     matter the most. For example, if you allowed compression
     or not in the material of your FEM model.

     If you are using Demons, on the other hand, the borders
     of the sphere will move to the *closest* border in the
     other sphere. That means that only displacement *perpendicular*
     to the border will be observed in the deformation field.
     Therefore two spheres that are displaced *will not* generate
     an horizontal displacement field. Instead, they will look
     like the squeezing/expanding case described above.



Please give it a try to the experiment suggested in (2)
(spheres at 1/10th of radius), and once you get a satisfactory
registration for that case, you can explore larger misregistrations,
and let us know what you find.



    Best regards,


         Luis



------------------------
Barbara Garita wrote:
> Hello Luis, I am sending this email to you because it bounces when I 
> send it to the itk user list email, probably because of the picture I've 
> attached, still I thought it was worth if you took a look at it.
>  
>  
>  
> 
> Thanks Luis, I've had a great time using Paraview and visualizing the =
>     displacement field result of the registrations I've ran!
>     I am in the process of making sense of the registrations fields, I've =
>     been registering simple gauss sphere of different scales and at =
>     different locations, however the vector fields don't look like I 
> expect =
>     them to look... look at the attached file, a simple case of =
>     translation...
>     1) the attached file is a simple translation, the red sphere is the =
>     fixed image, and the gray sphere is the moving image, I will expect 
> the =
>     vector field to be in one direction, from the fixed to the moving 
> image. =
>      Instead it goes along the y axis.  Any idea?
>     2) My question could be related to my lack of knowledge about the itk =
>     registration filter. Could you recommend any additional information on =
>     the registration filter, that could help me understand the results of =
>     the displacement field, and intelligently tweak the remaining 
> parameter =
>     (ei E, rho).
>     Thanks in advance!  Barbara
> 
>  
> 
> 
>   
>