[Insight-users] itkLinearInterpolateImageFunction

[Luis Ibanez]

Hi Anja,

What you describe, could be a potential way of implementing
a registration process, but may not be necessarily efficient.


You really should identify better the components of a registration
problem.



A) Transform:

    When you say "non-rigid" we assume that you are talking
    about a transformation that doesn't preserve angles.
    This includes : Affine and Similarity Transforms.



B) Interpolator:

    The role of the interpolator is to restore continuity
    on the moving image, so that you can query values in
    non-grid positions. This is needed regardless of the
    type of transform.

    Unless you are dealing with a Transform limited to
    translations parallel to the image axis moving at
    jumps that are multiple of the pixel spacing and, both
    of your images has the same direction, and their origins
    are separated by an integer multiple of the pixel spacing....



C) The Optimizer itself doesn't compute the Similarity Metric.
    That's indeed the role of the Metric classes. The Optimizer
    only chooses the location of the parametric space in which
    the Metric (Cost Function) must be evaluated. The computation
    itself is done by the Metric classes.



D) In the ITK implementation, there is no (explicit) resampling
    of the moving image at every stage of the registration.

    The reasons for not doing so are:

    1) It would require extra memory

    2) The image will degrade (lose high spatial frequencies)
       after passing through a series of multiple resamplings.


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


It seems that you are interested in understanding the
implementations of deformable registration in ITK.

There are three main versions:;


1) B-Spline Transforms (standard registration framework)
2) Dense Fields (Demons and its variants which are all PDE-based)
3) FEM (in which the image space is modeled as an elastic material).


Are you interested in particular in one of these specific methods ?



    Please let us know,


      Thanks


         Luis


-------------------
Anja Ende wrote:
> Hello,
> 
> Thanks guys. It really helps.
> 
> So, I am guessing when yoou are going non-rigid registration (say
> using free form deformation), then the registration algorithm would
> use something like B-splines to control the transformation of the
> input image. Then when optimizer computes the similarity matrix, it
> would resample the transformed image onto a regular grid and then run
> the similarity measure.
> 
> Is this a correct interpretation?
>