[Insight-developers] "itkSyNImageRegistrationMethod" cannot return its metric's transform properly

Sun Dec 2 17:15:16 EST 2012

Hi Nick and Hans,

I am sorry that my first email was not clear enough and thanks Hans for clarification.

I had done something like you presented in the previous email to get the current transform, and it was working; however, my focus was to get the current transform without making any changes to the ITK registration framework.

m_outputTransform is updated when optimization is completed, so I tried to get the current transform from the optimizer or metric, and I was thinking it would return its current transform like generic optimization framework, but it seems we have to do some light changes to itkSyNImageRegistrationMethod to finish this task.

Also, I think instead of updating the m_outputTransform at each iteration, we can add an interface to the optimizer framework of SyN filter to return the composer output if it is asked.

Thanks,
Ali

________________________________
From: Nicholas Tustison [ntustison at gmail.com]
Sent: Sunday, December 02, 2012 2:58 PM
To: Johnson, Hans J
Cc: Ghayoor, Ali; insight-developers at itk.org
Subject: Re: [Insight-developers] "itkSyNImageRegistrationMethod" cannot return its metric's transform properly

Yeah, that should be easy to do and I don't know that it would be a substantive
penalty---just unnecessary for minimum functionality.   It requires two instances of the
itkComposeDisplacementFieldsImageFilter.  If you take a look at the GenerateData()
function in itkSyNImageRegistrationMethod.hxx starting at line 537 and
going to line 550 you can see what I described in the last email to update
the output transform.

  typedef ComposeDisplacementFieldsImageFilter<DisplacementFieldType, DisplacementFieldType> ComposerType;

  typename ComposerType::Pointer composer = ComposerType::New();
  composer->SetDisplacementField( this->m_MovingToMiddleTransform->GetInverseDisplacementField() );
  composer->SetWarpingField( this->m_FixedToMiddleTransform->GetDisplacementField() );
  composer->Update();

  typename ComposerType::Pointer inverseComposer = ComposerType::New();
  inverseComposer->SetDisplacementField( this->m_FixedToMiddleTransform->GetInverseDisplacementField() );
  inverseComposer->SetWarpingField( this->m_MovingToMiddleTransform->GetDisplacementField() );
  inverseComposer->Update();

  this->m_OutputTransform->SetDisplacementField( composer->GetOutput() );
  this->m_OutputTransform->SetInverseDisplacementField( inverseComposer->GetOutput() );

We create two instances of the composer filter to create the forward and inverse
displacement fields and assign those to the output transform.  We would need to
take that snippet and copy that to wherever you wanted to query for the current
transform.  The iterative optimization all takes place in the StartOptimization()
function.  I'm assuming you could copy the snippet to somewhere around line
222 within a conditional where the metric value is assigned and the two transforms
to the middle have already been optimized for that iteration.

Nick

On Dec 2, 2012, at 3:31 PM, "Johnson, Hans J" <hans-johnson at uiowa.edu<mailto:hans-johnson at uiowa.edu>> wrote:

Nick,

We certainly do NOT want to update the m_OutputTransform at each iteration in the normal case, but we do need an option for getting the currentIterationTransform if requested (even with a potentially huge performance penalty).

Ultimately we need to make a movie of the deformation field at each iteration, and I don't care if I take a 20x speed penalty for making the movie.

If you could provide a hint of what needs to be changed (in ITK or ANTS) that would be great.

Thanks,
Hans

From: Nick Tustison <ntustison at gmail.com<mailto:ntustison at gmail.com>>
Date: Sunday, December 2, 2012 1:14 PM
To: Hans Johnson <hans-johnson at uiowa.edu<mailto:hans-johnson at uiowa.edu>>
Cc: Ali Ghayoor <ali-ghayoor at uiowa.edu<mailto:ali-ghayoor at uiowa.edu>>, ITK <insight-developers at itk.org<mailto:insight-developers at itk.org>>
Subject: Re: [Insight-developers] "itkSyNImageRegistrationMethod" cannot return its metric's transform properly

Okay, I figured it was just something I wasn't getting (particularly since
I was attempting an answer late last night).  I thought more about this
and I think I might have an idea of what the issue might be.

The SyNImageRegistrationMethod is used to optimize an itk
DisplacementFieldTransform since SyN doesn't fit in the generic
optimization framework.  The way optimization works is that we
initially create two additional displacement field transforms (in
addition to m_OutputTransform) called something like

fixedToMiddleTransform
movingToMiddleTransform

The first describes the forward and inverse transform from the fixed
image to the middle space.  The second describes the forward and
inverse transform from the moving image to the middle space.  During
the optimization, we strictly optimize just these two transforms.
m_OutputTransform only gets updated at the end of optimization by
composing the correct forward and inverse fields of the
fixed/movingToMiddleTransform(s).  We could potentially do this
composition at each iteration (i.e. update m_OutputTransform) but
that would be unnecessary computation (at least in terms of just getting
the final solution).

Nick

On Dec 2, 2012, at 11:34 AM, "Johnson, Hans J" <hans-johnson at uiowa.edu<mailto:hans-johnson at uiowa.edu>> wrote:

Ali,

You original e-mail is vague in its description of how you deterred that the displacement field has all zero's at every iteration.
-- Where you using a debugger?
-- How many zero values were listed?

Why do you need the const_cast in this code?  Could that change the behavior?

The use of const_cast is an indication that you are tyring to do something that is not supported.  Either ITK is deficient in not providing a const version of that member function, or you are trying to do something that has been explicitly prohibited.

Due to the ambiguity in your original posting, it is hard to figure out what your problem is.  You know that the the overall program produces a good result at the end, but you are trying to interrogate the process at every iteration, and for that you need to get the moving transform at each iteration.

***  I think that the problem you are having is exposing the current transform at each iteration to the observer.  ***

The response from Nick does not seem to address your real problem.

Please make a more thorough and very clearly stated problem, and then post enough information so that an external developer can replicate your error on their machine QUICKLY and EASILY.

Keep these facts in mind when requesting help from the ITK developer community:
1) We are almost all volunteers.  We help out of kindness not financial gain.
2)  If we can not replicate your error on my computer, we can not fix it.
3) Our time is valuable, if we need to guess, work, hard, or decipher code to determine how to replicate your problem, we will not have the time or energy even to get started
4) Provide both a) A description of the problem, b) description of how to replicate the problem, *** c) a description of how you believe the code should behave ***, and d) a proposed solution

Hans

PS: Nick Thanks for trying to address this request.  The original request was not very clear, and your effort showed that the end result is good (that was never in doubt).  The problem is interrogating for the current transform at every iteration.

From: Nick Tustison <ntustison at gmail.com<mailto:ntustison at gmail.com>>
Date: Saturday, December 1, 2012 8:24 PM
To: Ali Ghayoor <ali-ghayoor at uiowa.edu<mailto:ali-ghayoor at uiowa.edu>>
Cc: ITK <insight-developers at itk.org<mailto:insight-developers at itk.org>>
Subject: Re: [Insight-developers] "itkSyNImageRegistrationMethod" cannot return its metric's transform properly

Hi Ali,

I don't think that there's a deficiency.  I just ran the itkSyNImageRegistrationTest
and added the following snippet to write the displacement field to disk starting
at line 275.

  typedef itk::ImageFileWriter<DisplacementFieldType> WriterType2;
  typename WriterType2::Pointer writer2 = WriterType2::New();
  writer2->SetFileName( "/Users/ntustison/Desktop/testDisplacementField.nii.gz" );
  writer2->SetInput( outputTransform->GetDisplacementField() );
  writer2->Update();

The result was a non-zero displacement field that looked reasonable based
on the parameters specified by the test.

Nick

On Dec 1, 2012, at 8:30 PM, "Ghayoor, Ali" <ali-ghayoor at uiowa.edu<mailto:ali-ghayoor at uiowa.edu>> wrote:

Hello Developers,

In a registration framework, at each iteration, I get the current moving transform from the metric of the "itkSyNImageRegistrationMethod" filter; however, the returned moving transform has a displacement field with all zero values at every iteration.

To inspect this deficiency in ITK, I used "itkSyNImageRegistrationTest.cxx", and made some modifications in the test code for this purpose.
I added an iteration observer to the SyN registration stage:

    typedef CommandIterationUpdate<DisplacementFieldRegistrationType> DisplacementFieldCommandType;
    typename DisplacementFieldCommandType::Pointer DisplacementFieldObserver = DisplacementFieldCommandType::New();
    displacementFieldRegistration->AddObserver( itk::IterationEvent(), DisplacementFieldObserver );

Then, inside the observer class, I asked for the parameters of the moving transform at the current iteration:

   std::cout << " parameters: " << const_cast<TFilter *>( filter )->GetMetric()->GetMovingTransform()->GetParameters() << std::endl;
   std::cout << " fixed parameters " << const_cast<TFilter *>( filter )->GetMetric()->GetMovingTransform()->GetFixedParameters() << std::endl;

However, the Parameters of the moving transform is an all zero valued displacement field at each iteration.
Do you know why the registration filter does not return the right moving transform? Could you please help about this problem.

Thanks,
Ali
_______________________________________________
Powered by www.kitware.com<http://www.kitware.com/>

Visit other Kitware open-source projects at
http://www.kitware.com/opensource/opensource.html

Kitware offers ITK Training Courses, for more information visit:
http://kitware.com/products/protraining.php

Please keep messages on-topic and check the ITK FAQ at:
http://www.itk.org/Wiki/ITK_FAQ

Follow this link to subscribe/unsubscribe:
http://www.itk.org/mailman/listinfo/insight-developers

________________________________
Notice: This UI Health Care e-mail (including attachments) is covered by the Electronic Communications Privacy Act, 18 U.S.C. 2510-2521, is confidential and may be legally privileged.  If you are not the intended recipient, you are hereby notified that any retention, dissemination, distribution, or copying of this communication is strictly prohibited.  Please reply to the sender that you have received the message in error, then delete it.  Thank you.
________________________________

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.itk.org/pipermail/insight-developers/attachments/20121202/d3750a0b/attachment.htm>