[Insight-users] binary deformable registration : THERE IS NO SCIENCE WITHOUT REPRODUCIBILITY

[Luis Ibanez]

Hi Siddarth,


1) FEM methods when used in ITK for deformable registration are *NOT*
    physically based. That is, the real Young's modulus and Poisson's
    ratio of the tissues are totally irrelevant for the purpose of
    using ITK FEM-based deformable registration.

    What FEM does is to "simulate" the registration process as if the
    image was a piece of rubber. Note that the "forces" that ITK-FEM
    methods compute are based on derivatives of Image Metrics in blocks
    of pixels that surround the FEM grid nodes. Therefore the magnitude
    of this forces are *NOT* measured in Newtons, or any other physically
    consistent unit, but in the units of the Image Metric that you are
    using divided by units of pixel spacing.

    To put it simply:

              There is no reality in the FEM simulation.

    The values that you should explore for the Young modulus and Poisson
    ratio are based on the magnitude of the image metric derivatives.



1.1) You are right in that the Kappa cardinality is not available in
      the FEM framework.  Please add a feature request in the bug
      tracker:

             http://public.kitware.com/Bug/index.php


      In the meantime, please note,
      that mean squares will give you a very similar metric...



2,3) Reporting the methods, without the images, and without
      the parameters is quite useless.

      Parameters are an integral part of an algorithm. They
      make the difference between the algorithm working or not.

      Parameters are also dependent on the types of images that
      you feed into the registrations. Therefore, reporting
      parameters, without providing the images is useless as well.

      Again, this is the unfortunate result of years of subjecting
      our community to the decadence of the "Publish or Perish"
      mandate, where everybody is anxious for publishing "original"
      methods and nobody seems to be interested in "actually reproducing"
      the result reported by others.

      As a consequence we are accustomed to the simplistic
      and irresponsible claims of

                   "I tried method A and it didn't work"

      or

                   "Method A works better than Method B"

      such statements, are useless, and misleading, when they are
      not accompanied with a full report on how to repeat the
      experience:

            a) What image were used
            b) What parameters were set in the algorithms


The typical experience is that somebody will attempt to use
a method, try a couple of variations on the parameters, and
give up if results are not produced right away. The great
error is to jump from the technical statement:

       "Method A with set of parameters {P}, and image {I}
       produced the unsatisfactory results {R}"

to the simplistic

       "Method A does not work"


The first statement is science, the second one is superstition.



Please post your images (e.g. in a web site), along with the
parameters and the source code that you used for attempting
these registrations. Without that information it is very
difficult to provide you any serious advice.


    Thanks



       Luis


---


Let "Publish or Perish" perish.
Adopt "Publish Reproducibility or Perish".


=================================================================
Siddharth Vikal wrote:
> Hello ITK users,
> 
> I've been trying to do deformable registration of two segmented binary
> volumes (256x256x26). The volumes are first rigidly registered. Rigid
> registration achieves me more than 80% overlap. Now I need to improve
> further by correcting for shape deformation. I've tried the following:
> 
> 1) FEM based deformable registration, since I have rough idea of
> tissue properties (Young's modulus and Poisson's ratio) from pertinent
> literature. I chose mean squares metric as Kappa/Cardinality metric
> suitable for binary is not in current FEM framework; Further I chose 4
> pixels/element for mesh generation. The code runs and apparently
> converges but there is no real improvement in registration.
> 
> 2) Kappa + BSpline transform
> 
> 3) Cardinality + BSpline transform
> 
> I haven't had improvement in registration with any of above three
> methods. Besides, method (2) and (3) are extremely time-consuming
> (order of an hour)
> Now since the organ of interest is already segmented and
> rigid-registered, shouldn't it be easier to perform deformable
> registration?
> 
> Please advice me.
> 
> thanks,
> best regards,
> siddharth
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
> 
>