[Insight-developers] opinions on nonlinear warping functions

[Rupert Brooks]

Hi Richard,

This is a pretty open-ended question, but I will take a shot anyway.
I think its an interesting topic. My specific answer to your question
is no  - i've never compared the two approaches in practice either
generally or with specific implementations.  I used globally
parameterized registration throughout my thesis though, and did most
of my implementation in ITK.

If im not mistaken, the mni_autoreg operates in a block-based manner,
so i think the comparison is really between block-based approaches
(aka, iconic) and globally parameterized approaches.  While theres
many differences, i think some important ones are

- Block based approaches MAY get an advantage by optimizing over many
smaller subsets of parameters rather than a large global one.

- Block-based approaches are more complex to design and implement -
the question arises how to stitch together the blocks.  Piecewise
transformations may suffer from nonsmoothness, or even discontinuity
between the blocks.  Depending on the application, it might or might
not matter.  Expressing the transformation afterwards, and loading it
into other software might be tricky.  Application dependent.

- Block-based approaches permit the use of independent cost functions
per block.  For example, it can be advantageous to use a different MI
histogram per block, which can adapt to intensity variations over the
image.  I found this an advantage in a project i worked on. There may
be disadvantages to this too, fewer samples, etc.

- The specific case of B-splines has a number of limitations, however,
they are far from the only globally parameterized curved
transformation thats available.  I have used thin plate splines with
good results, but they do have certain problems.  Polyaffine
(unfortunately not implemented in ITK) would be worth looking at.

What experience i do have would suggest that for a practical
registration problems, the tuning and implementation of the algorithm
is extremely important.  So a kind of ho-hum algorithm which is well
tuned and implemented may outperform an excellent algorithm
haphazardly implemented.

One way to objectively investigate this question would be through a
registration grand challenge, or similar exercise.  (Theres a few
papers, I am thinking of one by Hellier et al on nonlinear
registration comparison, sorry, dont have the reference right here.)
Theres the RIRE project, but it is for rigid registration
only.(http://www.insight-journal.org/rire)  Something that looked
interesting popped up as an issue in the insight-journal a while ago
(http://www.insight-journal.org/browse/journal/16), but unfortunately,
it wasnt clear to me how it was going to work.  Anyone on the list
know of previous or ongoing open comparisons of nonlinear
registration?

Did you have a specific registration problem in mind?

Rupert B.

-- 
--------------------------------------------------------------
Rupert Brooks, Ph. D.

Attaché de recherches | Research Associate
Simulation des Matériaux Déformables | Simulation of Deformable Materials
Institut des matériaux industriels | Industrial Materials Institute
Conseil national de recherches Canada | National Research Council Canada
75, de Mortagne, Boucherville, Québec, Canada, J4B 6Y4
Gouvernement du Canada | Government of Canada

On Mon, Apr 20, 2009 at 9:34 PM, Richard Beare <richard.beare at gmail.com> wrote:
> Hi,
>
> Does anyone have experience/opinions on the benefits and tradeoffs
> using a hierarchy of linear transforms to represent a nonlinear warp
> (as per mni_autoreg) as opposed to a direct nonlinear representation -
> bsplines etc?
>
> Thanks
> _______________________________________________
> Powered by www.kitware.com
>
> Visit other Kitware open-source projects at http://www.kitware.com/opensource/opensource.html
>
> 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
>