[Insight-users] Re: reg similarity measure(deformable registration)
luis.ibanez at kitware.com
Sun Jul 11 14:14:12 EDT 2004
Please read the Tutorials in Image Registration available at
You should also read the chapter on Image Registration in
the ITK SoftwareGuide
About your questions:
You can use deformable registration without landmarks.
As a matter of fact, none of the deformable registration
methods that you listed require landmarks to be defined.
You can optionally use landmarks in FEM if you feel
inclined to do so.
2) "What method should give better results" ?
That's a retoric question that only make sense in the
"paper-publishing" context of "my method is better than
yours". That's only useful in the Matrix...
In real life there is no such a thing as a "better method",
since the performance of each method is extremly dependent
on the parameters that you select for it. What you will find
in real life is the ubiquitous trade-off of quality versus cost.
Where cost involves computation time, memory allocation, user-time
required for fine tunning parameters, and of course, amount of
liters of coffee to be consumed in the process.
So, turn the question around and ask yourself first:
- How many weeks do I have to get this registration done ?
- How long should a registration computation take ?
- How many weekends and sleepless nights am I planning
to pass fine tunning the parameters of the registration ?
- How many journal and conference papers am I willing to
study in order to get the background needed for getting
the best out of this method ?
- How much registration precision is required for the
application for which this registration is intended ?
Your choice of Deformable Registration method will be
quite straight forward once you answer those questions :-)
Here are some guidelines
that you may want to consider at that point:
Demons will be appropriate if the deformations in the
image are small. Where "small" is defined as "being in
the range of the transition regions of the anatomical
structures in the image". That is, if the transition in
intensity between one tissue and another in your image
is 3 pixels wide,... then Demons will be limited to
correct for deformation equal or smaller than 3 pixels.
You can of course overcome this limitation by using
Demons in a Multi-Resolution context.
There are few parameters in Demons, but the
computational load is high.
BSplines have also few parameters to fit, mostly those of
FEM is the most flexible method, probably the one from
which you can get better quality results, but also the
one that will require more effort on your part to get
it working correctly in your images.
I would suggest you to start with BSplines, using a small
grid (e.g. 8x8x8), and if needed move to a larger grid.
Using a larger grid will be needed only if your images
have local deformations that go in different directions.
3) You are right in the first statement and wrong in
Mutual-Information have prove to be appropriate for
ITK *does* provide implementations of the MutualInformation
There are actually four different implementation
of Mutual Information in ITK
- Viola Wells
- Mattes et al.
- Histogram based
4) You will find the full list of image metrics available
in ITK in the Doxygen documentation.
As you can see there are twelve different Image-to-Image
Metrics available in ITK at this point.
You are certainly welcome to write new implementations of
Image metrics. If you decide to do so, we strongly encourage
you to contribute your code back to ITK so other users can
take advantage of your efforts.
Dalal, Dhaval wrote:
> hey luis
> I know that ITK provides multiple methods for performing deformable
> In particular:
> 1) BSpline Transform
> 2) Demons deformable registration
> 3) FEM-Based deformable registration
> I want to register 2 MRI datasets, their shapes are quite different
> hence i will use deformabale registration method.
> my questions are
> 1) I do not have any fixed landmarks, is it possible to register without
> known landmarks in these 3 methods.
> 2) which of these 3 methods should give the better result.
> 3) Regarding the similarity metric, i think it is an accepted fact that
> mutual information gives the better result for multi-modal registration.
> itk i do not think provides a framework for mutual information based
> registration? why is that so.
> 4) other similarity measures assume the intensity to be the same in both the
> images. not mutual information.
> what are the similarity measures implemented in ITK.
> is it possible to write the code for mutual information.
> Kindly help me out please.
> your help will be greatly appreciated.
> Dhaval Dalal
More information about the Insight-users