[Insight-users] ConfidenceImageFilter

[Luis Ibanez]

Hi Samuel,

The ConfidenceImageFilter is surprisingly efficient
for segmentation despite its apparent simplicity.

It has been quite succesful in MRI and CT images
both for segmenting normal anatomy and for segmenting
globular tumors.

You may get better results by first smoothing the
images with any of the anisotropic diffusion filters
that provide edge-preserving smoothing.

for example:

- CurvatureFlow
- CurvatureAnisotropicDiffusion
- GradientAnisotropicDiffusion

http://www.itk.org/Insight/Doxygen/html/classitk_1_1CurvatureFlowImageFilter.html
http://www.itk.org/Insight/Doxygen/html/classitk_1_1CurvatureAnisotropicDiffusionImageFilter.html
http://www.itk.org/Insight/Doxygen/html/classitk_1_1GradientAnisotropicDiffusionImageFilter.html

pre-smoothed images simplify the task of the
ConfidenceConnectedImage filter since they provide
regions with more specific statistical distributions.

A demo application using these filters is availabe
under the Insight/Applications directory:

       RegionGrowingSegmentation

----

The Mean and Variance around the seed are computed in the
first iteration of the filter by using the following helper
classes:

itk::MeanImageFunction
itk::VarianceImageFunction

http://www.itk.org/Insight/Doxygen/html/classitk_1_1MeanImageFunction.html
http://www.itk.org/Insight/Doxygen/html/classitk_1_1VarianceImageFunction.html

These two functions use the SmartNeighborhood iterator and
define the size of the neigborhood to a radius of 1. That
is a 3x3 classical neighborhood in 2D and a 3x3x3 in 3D.

So basically, the data over which the initial mean and variance
are computed in 3D are the 26 closest pixels around the seed and
the seed itself.

Now that you raise this question, it may be interesting to increase
the size of the initial neigborhood in order to improve the chances
of getting a more representative estimation of the region statistics.

This will reduce the sensitivity of the algorithm to the selection
of the seed point.

Probably we may want to add a method:

          void SetInitialNeigborhoodRadius(int)

to the itk::ConfidenceConnectedImageFilter.


----


Please let us know if you have further questions.

Thanks

    Luis


=======================================
Samuel Rodríguez Bescos wrote:
> Hello everyBody,
> 
>  
> 
> Could annybody tell me where I can find information about the algorithms 
> using in the implementation of the filter ConfidenceImageFilter and for 
> wich type of images have best results?.
> 
>  
> 
> I read the information at the website but I have a question. How are the 
> pixels at the neighborhood seed evualuated in the first iteration?. What 
> are the mean an variance values?.
> 
>  
> 
> Thanks in advance,
> 
>  
> 
> Sam
>