[Insight-users] registration consistency

[Luis Ibanez]

Hi Qing,

The class

    itkMattesMutualInformationImageToImageMetric
http://www.itk.org/Insight/Doxygen/html/classitk_1_1MattesMutualInformationImageToImageMetric.html

uses the iterator

    itkImageRandomConstIteratorWithIndex
http://www.itk.org/Insight/Doxygen/html/classitk_1_1ImageRandomConstIteratorWithIndex.html

for generating a random sample of pixels from the Fixed image.

The ImageRandomConstIteratorWithIndex generates the index
position of the pixels by making calls to the class

    itkMersenneTwisterRandomVariateGenerator
http://www.itk.org/Insight/Doxygen/html/classitk_1_1Statistics_1_1MersenneTwisterRandomVariateGenerator.html

This random number generator is instantiated as a *Singleton*
in your program.


Upon construction its seed is initialized to the value "121212".
(See line 541) of itkMersenneTwisterRandomVariateGenerator.h
in Insight/Code/Common.

A call to its "SetSeed()" method [with no arguments] will use a
new seed computed from a hash of the current time as it is returned
by vcl_time() and vcl_clock().


--


As a user of the registration framework, if you want to guarantee
reproducible results when using the Mattes MI metric, you will have
to initialize the seed to a given value (any value, as long as you
always use the same).


The Mattes MI Metric offers the method:

            metric->ReinitializeSeed(int seed)


If you call this method with the *SAME* seed value, *BEFORE* you
call the StartRegistration() or the Update() method of the
itkImageRegistrationMethod  object, then your registration
results should be reproducible, regardless of whether you are
registering other images before.

To be more specific, *INSIDE* your for loop, you must call Reinitialized
Seed before triggering the registration process for the current dataset.



    Please note the in its current form, the MersenneTwister random
    generator *IS NOT THREAD SAFE*. A bug on this issue is pending to
    be solved.


    Please *DO NOT* use the GetInstance() method of the random number
    generator, as this API is likely to be removed from the toolkit,
    in order to solve the lack of thread-safety in this random number
    generator.


---


Please make sure *also* that you are initializing the Transform inside
the for loop, so that each image is starting always with the same
transform. You may be doing this by initializing the Transform,
or by using an array of InitialParameters for the
ImageRegistationMethod. In either case, you must reset it at every
iteration of the for loop.



       Regards,


         Luis




-------------
qing xu wrote:
> Hi,
> 
> Does the same registration code with the same input always give the same 
> output?
> 
> I am doing a 3D registration for a dataset consisting of 50 volumes. I 
> put the registration code in a loop so that the registration can be done 
> one by one for each volume. However, the reults are much different from 
> what I got by processing one volume once a time. The only reason I can 
> think of is that running reigstration for too long would consume too 
> much memory, so the registration performance is affected. Is that 
> possible?  Does anybody have the same experience? Thanks in advance.
> 
> 
> (my metric is mattes mutual information and transform is 
> VersorRigid3DTransform)
> 
> 
> Qing
> 
> _________________________________________________________________
> http://newlivehotmail.com
> 
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
>