[Insight-users] some information on the registration example
methods
Luis Ibanez
luis.ibanez at kitware.com
Mon, 15 Mar 2004 22:55:04 -0500
Hi Michael,
The reason why Multi-Resolution is faster is that
you solve the problem progressively by first computing
large deformations from reduced versions of the original
images. If you even need to go faster, a good option
is to initialize the deformation field using landmarks.
You will find an example on this in
Insight/Examples/Registration
It is not obvious to figure out a mechanism for 'driving'
the registration process from the Observer. Certainly it
is technically possible, but the challenge is that you
would have to analyze the deformation field and figure out
a parameter tweaking capable of improving the next iterations
of the solver. Sounds like a good subject for a PhD dissertation,
or at least good topic for a journal paper :-)
--
The issues of FixedArray versus Array as carriers of
parameters in the registration methods have been revisited
recently.
This doesn't affect the Demons algorithm since this one
is not solved using typical ITK optimizers but by using
a PDE solver.
Regards,
Luis
------------------
Michael G wrote:
> Hi Luis,
>
> Thank you for all the information - I've been trying different
> approaches as you suggested and they have resulted in reduced
> computation times.
>
> I still have a question about the
> MultiResolutionPDEDeformableRegistration template.. In the
> MultiResolutionImageRegistrationMethod template there are methods to set
> the optimizer at the start time and through an observer. I can't figure
> out how to use an observer to 'push' the
> MultiResolutionPDEDeformableRegistration template in the right direction
> during registration - is it not possible, and the reason this method
> will be faster than using the pure DemonsRegistration is that you can
> set the number of levels and it will work on those? (instead of just the
> original) Are there any tweaks you can do to make it faster?
>
> Lastly, while looking for other posts about registration I came across a
> series of posts having to do with itk::FixedArray<> and the itk::Array<>
> in regards to speeding up some registration methods. The posts were from
> last May, did anything come of it?
> http://www.itk.org/pipermail/insight-users/2002-May/000427.html
>
> Thanks for your help Luis, if all the above is impossible it's not a
> problem, I'm just curious - this toolkit is amazing,
>
> Thanks,
> Michael
>
>
>> From: Luis Ibanez <luis.ibanez at kitware.com>
>> To: Michael G <michaelguiliametti at hotmail.com>
>> CC: insight-users at itk.org
>> Subject: Re: [Insight-users] some information on the registration
>> example methods
>> Date: Wed, 03 Mar 2004 20:19:05 -0500
>>
>>
>> Hi Michael,
>>
>>
>> It is great to have users who
>> post messages when things work ! :-)
>>
>>
>> --
>>
>>
>> About the deformable registration example,
>> You can reduce the computation time by
>> using a Multi-Resolution approach. This
>> particular example is using the Demons
>> algorithm and was implemented by Lydia Ng
>> at Insightful. The DemonsRegistrationFilter
>> derives from the generic class:
>>
>> PDEDeformableRegistrationFilter
>> http://www.itk.org/Insight/Doxygen/html/classitk_1_1PDEDeformableRegistrationFilter.html
>>
>>
>>
>> The MultiResolution variant of this class is the
>> http://www.itk.org/Insight/Doxygen/html/classitk_1_1MultiResolutionPDEDeformableRegistration.html
>>
>>
>> What you want to do in this case is to progressively
>> refine the deformation field by computing first a
>> low resolution version of it. In this way you reduce
>> the time to convergence and improve the robustness
>> of the registration.
>>
>>
>> You will find information on the ITK registration
>> framework in the Software Guide:
>>
>> http://www.itk.org/ItkSoftwareGuide.pdf
>>
>> In particular Chapter 8.
>>
>> (you can order the paper version of this document
>> from Amazon.com and from Kitware's estore)
>> http://www.kitware.com/products/itkguide.html
>>
>>
>> You will also find details on the theory behind
>> algorithms implemented in ITK, in the book
>> "Insight into Images" published by A.K. Peters
>> http://www.akpeters.com/book.asp?bID=206
>>
>>
>> The tutorial session on deformable registration
>> presented by Dr. James Gee will also be illuminating:
>> http://www.itk.org/CourseWare/Training/NonRigidRegistrationMethods.pdf
>>
>>
>>
>>
>> Please let us know if you have further questions.
>>
>>
>>
>> Thanks,
>>
>>
>>
>> Luis
>>
>>
>>
>> ------------------------------
>> Michael G wrote:
>>
>>> Hello,
>>>
>>> I am using the technique in ImageRegistration1 to motion correct a
>>> set of dicom images and it's working great. It takes about 3-4
>>> seconds per image.
>>>
>>> I noticed the other great examples, specifically the
>>> DeformableRegistration2, because it seems to offer even cleaner
>>> results. This method is taking a great deal longer to run per slice,
>>> around 11 seconds on the same machine.
>>>
>>> Are there any methods to improving the speed for these methods? Also
>>> are there more resources on the science behind these registration
>>> methods that itk has provided for us? I could only find
>>> http://www.itk.org/CourseWare/Training/RegistrationMethodsOverview.pdf
>>>
>>> Thank you very much for any info,
>>> Michael
>>>
>>
>>
>>
>> _______________________________________________
>> Insight-users mailing list
>> Insight-users at itk.org
>> http://www.itk.org/mailman/listinfo/insight-users
>
>
> _________________________________________________________________
> Fast. Reliable. Get MSN 9 Dial-up - 3 months for the price of 1!
> (Limited-time Offer)
> http://click.atdmt.com/AVE/go/onm00200361ave/direct/01/
>
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
>