[Insight-users] registration help needed

[Luis Ibanez]

Hi Soumen,



A) Mutual Information is not a determininstic metric.

This means that experiment run in different machines *will not*
produce exactly the same results.  The reason is that MI samples
random pixels from the image in order to estimate the joint pro-
bability distribution.  The selection for pixels may be different
on different machines, even on different runs in the same machine.

Another factor that may affect your results is the fact of compiling
your code for Debug/Release/RelWithDebInfor... and so on.  The type
of numberical registers used on each case, will be different and
therefore the cumulation of errors over time will lead you to different
results.

In any case, if your program is converging in some platforms and failing
in others, it simply means that your parameters are set up too close to
critical values.  You may want to use smaller learning rates for the
optimizer, and more conservative values for the translation scaling
in the Transform parameters.




B) PET-CT cannot always be registered successfully.

It all depends on what kind of isotope was used in the PET scan
and what anatomical structure was targeted. In many cases, PET scans
lack any anatomical information that may make possible to register them
to a CT scan.  So, the first question here is:

Do you see in the PET scan any anatomical structures that can
unequivocally be associated to structures in the CT ?




C) Generalizing the registration process is not what you want.

Generality, is actually what you already have in ITK.
What you need is specialization, which is the opposite.
You need a specialized version of registration that will
be customized for the characteristics of PET-CT registration.

The way to get there is by running an study in a set of PET-CT
images and finding parameters that work for all the pairs on the
set. Then you hard-code those parameters and encapsulate the
whole in a single class with a name such as itkCTtoPETRegistration.




D) If you are interested in exploring new registration techniques,
    I thinks that your previous question just raised the point that
    most users will like to get solved:

                        "Parameter setting"

    this is no-doubt, the most painful aspect of performing registration,
    (as well as segmentation).

    You may want to explore techniques for making self-adapting
    registration methods.  In such methods, you may want to make the
    optimizer and the metrics aware of the ranges of values that
    you expect, as well as their variations. This will allow a
    controller class to identify when the optimizer is advancing
    too much or too little during the registration and therefore
    it may be able to take a corrective action.





   Regards,



       Luis





---------------------------------
soumend at magnum.barc.ernet.in wrote:

>  HI!!!
>  
>  I am facing one peculiar problem in doing MI registration. My code
>  registeres
>  one CT dataset(3D) with a PET dataset(3D). My problem is the registration
>  problem do get completed in WINXP(pro) but in other windows platform like
>  win2k,win98 it is giving exception originating from the fact that most of
>  the
>  point are mapped outside. I tried the same code in RH-LINUX 7.3 and 9 and
>  got
>  the same exeception. Why this is happenning.......I couldnt find out..plz
>  help
>  me out.
> 
>  I have another query. The result i am getting with CT-PET in WINXP is not
>  satisfactory. The images are visibly shifted from one another. I have used
>  following parameters.........
>  no. of bins - 200, no. of samples - 10000 and
>  Optimizer scales [1,1,1,1,1/750,1/750,1/750,1/750] as fixed image dimension
>  is
>  512X512.
>  Optimizer max and min step lengths are 0.1,0.001 resp.
>  
>  I have tried with diffrent value of these parameters but finally settled in
>  this
>  set of value. If I want to improve upon more how should I proceed. In this
>  regard I one to ask one thing is it feasible to generialize the registration
>  process such that user dont have to interact and set parameters for every
>  pair
>  of image instead the s/w will take care of that.
>  
>  Lastly I need a suggesion ....If I want to develop a new algo for
>  registration
>  how sholud I proceed...as I have started reading Information theory and
>  satistics a bit and the original paper of Violla. What I mean to ask that
>  what
>  are the features people look for a good ,fast and robust registration
>  process.
>  Probably I have not been able to frame the question but that exactly why I
>  am
>  asking the question.
>  
>  Thanks for ur patient reading.....hope u will help me out.
>  
>  Thanx and regards.
>  
>  Soumen Dey 
>  
> 
> 
> -------------------------------------------------
> 
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
> 
>