Hi Markus,
Thanks for your clarification.
I see better now the problem that you were pondering.
The short answer is :
Yes, you could do this with ITK,
The long answer is:
Yes, as you already anticipated, even though this is feasible,
it is also computationally very expensive.
The approach that you suggested is a soft of generalization
of the Demons class for any type of ImageMetric, not just
Mean Squares. It is possible to implement such combination,
but it will require a massive amount of computation.
Note that this is the reason why BSplines and FEM methods
are used, as a mechanism for reducing the parameter space
from the millions of pixels in a deformation field, to the
hundred or thousands of BSPline coefficients, or FEM node
displacements.
Even with the BSpline casses, the LBFGSB algorithms gets
already streched thin.
The scenario will be interesting however as a challenge for
a use of massive parallel approach.
It has the advantage, as you pointed out, that you could
define many different regularization methods, and will
provide an interesting testing framework for new ideas.
Combining this with a multi-resolution approach for
computing the deformation field, will probably be a
convenient mix.
Regards,
Luis
---------------------
Markus Weigert wrote:
> Hi Luis,
>
> thanks for your reply, but at what I was really targeting with my
> question is
> the following (more theoretical):
> One could consider the transformation field (in any dimension, say 3D)
> between two images
> in terms of a generic transform and thus optimize the metric between
> the two images by
> purely optimizing the deformation field entries (starting with all
> deformation values set to 0).
> Here, one may use LBFGSB optimizer or any optimizer from the toolkit to
> optimize the metric
> with respect to all the deformation vectors entries.
> In this case, the metric class would need a deformation field to be
> plugged in and need to return
> the metric value and derivatives with respect to the deformation field
> entries.
> Of course, the parameter spaces with which one had to deal in this case
> may reach up to hundreds
> of millions of unknowns to be optimized for large 3D datasets (at
> highest resolution).
> Also, one could use all kinds of regularisation on the calculated
> deformation field by "simply"
> optimizing a cost function taking metric and regularisation term
> (elastic, fluid, whatever) into account.
> Do you think that this naive approach would be feasable using standard
> optimization routines like LBFGSB
> as it is implemented in the toolkit or is it just not feasable?
>
> Kind regards,
> Markus
>
>
>
> ----- Original Message ----- From: "Luis Ibanez" <luis.ibanez at kitware.com>
> To: "Markus Weigert" <m.weigert at fz-juelich.de>
> Cc: <insight-users at itk.org>
> Sent: Tuesday, September 26, 2006 12:52 AM
> Subject: Re: [Insight-users] Question on image to image metric
> implementations
>
>
>>
>> Hi Markus,
>>
>> Please read the ITK Software Guide
>>
>> http://www.itk.org/ItkSoftwareGuide.pdf
>>
>> in particular the Image Registration Chapter,
>> and the section on "Deformable Registration".
>>
>> You can use a BSplineDeformableTransform as the
>> spatial transformation in the Image Registration
>> Framework. From the resulting BSpline you can
>> obtain a deformation fields.
>>
>> You will find coding examples in that chapter,
>> and the source code of the example is available
>> at
>>
>>
>> Insight/Examples/Registration
>>
>>
>> You may also be interested in the Demons registration
>> filter and the FEMRegistrationFilter.
>>
>>
>> Regards,
>>
>>
>>
>> Luis
>>
>>
>>
>> ========================
>> Markus Weigert wrote:
>>
>>> Hi all,
>>> when using the ImageToImage metrics (for example
>>> itk::MutualInformationImageToImageMetric), one needs to provide
>>> a spatial transformation, which is used by the metric (for example to
>>> compute derivatives).
>>> My question is what about using a deformation field (WarpImageFilter)
>>> as "transformation",
>>> what may also be seen as the more general approach???
>>> One could consider the deformation field as a transformation as well
>>> (with much more parameters, of course)
>>> and calculate derivatives with respect to the deformation values (to
>>> control for example some registration process).
>>> Is the framework open for this and how expensive would it be to
>>> implement such an approach (not only complexity
>>> of calculation but also effort of implementation)?
>>> Thanks for any suggestions,
>>>
>>> arkus
>>> ------------------------------------------------------------------------
>>>
>>> _______________________________________________
>>> Insight-users mailing list
>>> Insight-users at itk.org
>>> http://www.itk.org/mailman/listinfo/insight-users
>>
>>
>
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
>
>