[Insight-users] Speed Image for Level Set Filters

[Luis Ibanez]

Hi Neha,

The purpose of the speed image is to make the level set
slow down close to the edges of anatomical structures,
and to make it go faster on the internal regions.

Therefore you want the speed image to have low
(dark) values on the pixels close to edges of
anatomical structures and to have high (bright)
values in the middle.

This is explained in detail on the LevelSets chapter
of the ITK Software Guide.

    http://www.itk.org/ItkSoftwareGuide.pdf



If you use the Gradient magnitude directly as
a speed image, you will get the opposite effect.
You will make the level set to move fast across
edges, and to slow down inside the structures.

That's the reason why you need to revert the
intensities of the Gradient Magnitude image.



Note that you don't have to use the Sigmoid.
Any other filter that reverts the image intensities
will work here.

A filter that is commonly used in the LevelSet
literature is the negative exponential

http://www.itk.org/Insight/Doxygen/html/classitk_1_1ExpNegativeImageFilter.html

You can also use the EdgePotential image filter.

http://www.itk.org/Insight/Doxygen/html/classitk_1_1EdgePotentialImageFilter.html



Note that the choice of the filter is critical, since
the speed image (or feature image) is the most important
information used for controlling the evolution of the
LevelSet.  The segmentations that you obtain from LevelSets
are as good or as bad as your feature image.



    Regards,


       Luis




-------------------
neha k wrote:

> Hello,
> 
> For Shape Detection Level Set Filter and other Level Set filters, speed 
> image is obtained from Gradient Magnitude.  Then the speed is calculated 
> from Speed Image which is obtained from Sigmoid filter.  How is speed 
> calculated from it?  It looks like output of sigmoid is obtained by 
> inverting gradient magnitude ouput. Is it possible to use gradient 
> magnitude output as a speed image? I am not clear on real advantages and 
> purpose of using sigmoid filter.   Any input of this is appreciated.
> 
> Thanks,
> 
> Neha
> 
>  
> 
>