[Insight-users] Re: about demons algorithm

[Luis Ibanez]

Hi Guorong,

The Demons algorithm implemented in ITK uses the concept
of optical-flow in order to compute forces that are then
used for generating a deformation field.

The basic assumptions of optical-flow is that the images
are of the same modality and that they have almost identical
intensity distributions.

The process of Histogram Matching that you see in the
example DeformableRegistration2.cxx is there only in case
the two images to be registered don't quite have the same
intensity distributions. In that case, the histogram matching
filter will compensate for the differences and bring one
image to have a closer intensity distribution to the other.

If you image already have very close intensity distributions,
then you don't need the Histogram Matching preprocessing, and
can simply remove it from the pipeline.

If you find that the registration works better without the
Histogram Matching, then that's enough reason for removing
that pre-processing step.


Using multi-resolution is desirable almost for all image
registration problems. It is not surprising that if you
use multi-resolution you got faster and more accurate results
that with the example where no multi-resolution is done.

The issue of validation is a never ending problem, basically
because there is no such thing as "reality" in image processing.
If you want to attempt to perform validation, or evaluation
of deformable registration algorithms, one good way to go is
to artificially induce deformations in an image, and then
attempt to register it back with its original using the algorithms
that you want to compare.

You will find among the ITK examples, two programs that can be
used for artificially deforming images. They are:

   Insight/Examples/Registration/
                       LandmarkWarping2.cxx
                       DeformationFieldInitialization.cxx


One uses KernalSplines and the other uses BSplines.
Since you fabricate the deformation field by yourself,
then you have a refence against which you can compare
the results of the registration method.


Note that deformable registration may have many different
solutions. You may have to consider other factors such
as whether the resulting deformation should be incompressible
or not. For example, Demons, and BSpline transforms do not
preserve the volume of tissues.  FEM on the other hand may
be configured for ensuring such conservation.



BTW,
Would you like to contribute your implementation of Demons to ITK ?


If you find that it works better than the current implementation,
it will certainly be useful to other users. Please let us know,
and we will guide you to the process of adapting your code to
ITK standards.



    Regards,



       Luis



---------------------
Guorong Wu wrote:

> hello luis:
> 	Sorry to mail to you directly because my mail is too big . 
> 	I am studying "demons" algorithm. In example "DeformableRegistration2.cxx" the program does histogram matching before DemonRegistration. But I find the result is better without histogram matching. As you can see in the attachment files, axial.png is the fixed image, axial-rot05.png is just the 5 degree rotation of fixedimage. And result4.png is result using demons together with histogram matching, result3.png is the result without using histogram matching. 
> 	My first question is whether it is nessary to do histogram matching. 
> 	At the same time, I make my own demons program using multi-resolution. To regist axial.png and axial-rot05.png, DeformableRegistration2 take much longer than mine. rusult1.png is the result in my program without using histogram matching and result2.png is using histogram matching. It is obvious result1.png is better than result2.png. But the SSD(sum of squared diffence) per pixel of result1 is 58.04, while result2 is 49.84. So my second question is how to explain or how to evaluate two non-rigid registration algorithms.
> 	Thanks. Good luck! 
>  				
> 
> 　　　　　　　　Guorong Wu
> 　　　　　　　　grwu at sjtu.edu.cn
> 　　　　　　　　　　2004-07-06
> 
> =============================================================
> Guorong Wu	
> The Department of Computer Science and Technology
> Shanghai Jiao Tong University
> 1954 Huashan RD
> Shanghai China    
> 200030
> Tel:  +86 21 3226 1236  , 6293 2089
> =============================================================
> 
> 
> ------------------------------------------------------------------------
>