[Insight-users] 3D inter-subject nonlinear image registration (same modality)

[Luis Ibanez]

Hi Jessie,



0) I assume that by "non-linear" registration you
    actually meant "non-rigid" registration.



1) There is no such a thing as the "Best" method.

    The concept of the "best" method is only useful for
    publishing papers and it is devoid of any technical sense.

    Whoever claims to have the "best" method for doing
    something is just trying to sell you "his/her" method
    and is failing to tell you the entire story.


    In practice you will find "suitable" methods, and each
    one will have trade-offs that you will have to evaluate
    in order to find a balance between how much computation
    time you can afford, what kind of deformations do you want
    to allow, and how much precision do you expect in the final
    fitting.



    Your options for performing non-rigid registraiton in ITK
    are the following:


     1) Demons
     2) FEM-Based
     3) BSplines
     4) LevelSet


     For details on each one you may want to look respectively at


     1) Demons:

        Insight/Examples/Registration/
                 DeformableRegistration2.cxx
                 DeformableRegistration3.cxx

         and the ITK Software Guide

             http://www.itk.org/ItkSoftwareGuide.pdf

         Section 8.13, pdf-page 323



      2) FEM-Based

         ITK Software Guide

             http://www.itk.org/ItkSoftwareGuide.pdf

         Section 8.12, pdf-page 318

         and the Tutorials  http://www.itk.org/HTML/Tutorials.htm

         in particular the Session "Non-Rigid Registration"

http://www.itk.org/CourseWare/Training/NonRigidRegistrationMethods.pdf



      3) BSplines

            Insight/Examples/Registration/
                        DeformableRegistration4.cxx
                        DeformableRegistration6.cxx
                        DeformableRegistration7.cxx
                        DeformableRegistration8.cxx



       4) LevelSet deformable registration

http://www.itk.org/Insight/Doxygen/html/classitk_1_1LevelSetMotionRegistrationFilter.html

            Insight/Examples/Registration/
                        DeformableRegistration5.cxx







Before you engage in your exploration of this methods you MUST
ask yourself the following two questions:


A) What do you want to keep unchanged during the registration.
B) What do you want to allow to change during the registration.


The answers to these two questions will determine what methods
are suitable for your medical image registration problem.







3)  A non-rigid registration method for brain images of size
     256 x 256 x 200 can typically take one hour of execution
     time in a modern desktop machine (e.g. Pentium 4 at 2Ghz).

     computation time can be reduced dramatically if you facilitate
     a good initialization, by allowing the user to provide
     landmarks, or by performing an Affine registration before
     starting the "non-rigid" one.

     computation time can also go up to several hours depending
     on the method that you select and the convergence criterion
     that you choose.



     "Accuracy" is ill-defined in the context of medical image
     registration. In particular when you are registering the
     brains of two different subjects because the anatomical
     similarities between two brains are limited to the large
     scale structures.  Strictly speaking, you can only define
     accuracy on the large structure, since those are the only
     ones that can be considered homologous between the two
     brains.

     Once you start looking at the details of the Sulci and Gyra,
     the structure of each brain is different. Therefore you can
     only match to different brains by "un-naturally" forcing the
     foldings of one brain to match the foldings of another.

     So,... in practice, you can get as much "Accuracy" as you want,
     at the price of computation time and un-natural deformation
     fields....


     The question is:

          "How much Accuracy do you want ?"

          "Where in the brain you need more accuracy ?"

          "How much time do you have available ?"



You may want to take a new look at the specific medical problem that
you are trying to solve and get a clearer definition of the purpose
and constrains of your problem.


E.g.
do the brains have pathologies ?  (like tumors)

are the ages of the subjects similar ?
(adolescent brains are quite different from adult brains)

do the subjects have the same gender ?






    Please let us know if you have further questions,




       Thanks



          Luis




-------------------------
(Jessie) Ting Guo wrote:

> Hello everyone,
>  
> I have some question regarding to using ITK to do 3D inter-subject 
> nonlinear image registration. Would anyone who has some experiences with 
> this give me a hand?
> 
> 1. Which method is the best in ITK to do 3D inter-subject nonlinear 
> image registration? Which paper was used to develop this method?
> 
> 2. Approximately, how fast and accurate can this method to register 
> an image to another, such as registering a full brain image to another 
> one? (please also specify the hardware platform.)
>  
> Thank you very much!
> Jessie
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users