[Insight-users] leaky segmentation

[Luis Ibanez]

Hi Toon,

If you zoom into the image and look closely at the
separation between the tubes. How many pixels there
are in the strip separating the tubes ?

If you find at least four pixels, then you have
some hope of using derivatives in order to separate
the two structures. Therefore you could use
GeodesicActiveContours and increase the scaling of
the Advection term.

Advection is computed using a vector field that
points tot the edges of structures in the image.

---

If you find less than four pixels of separation,
then you are in trouble and will have to be more
creative...     :-/
The typical option will be to use a model-based
segmentation in which you introduce prior knowledge
about the shape.

You may want to look at the following classes:


A) ShapePriorSegmentationLeveSetImageFilter

http://www.itk.org/Insight/Doxygen/html/classitk_1_1ShapePriorSegmentationLevelSetImageFilter.html

B) Its derived class GeodesicActiveContourShapePrior

http://www.itk.org/Insight/Doxygen/html/classitk_1_1GeodesicActiveContourShapePriorLevelSetImageFilter.html

   From the paper: Leventon, M.E. et al.
   "Statistical Shape Influence in Geodesic Active Contours", CVPR 2000.
   http://www.ai.mit.edu/people/leventon/Research/0006-CVPR/cvpr00.pdf


---

For your suggestion of two competitive snakes, this
equivalent to implement the Multi-phase Level Sets,
for example:
http://portal.acm.org/citation.cfm?id=598537&dl=ACM&coll=portal

Currently ITK doesn't have such methods available,
but it shouldn't take much modifications on the
LevelSet framework in order to implement those
methods.




Regards,



    Luis


----------------------
Toon Huysmans wrote:

> Hi,
> 
>  
> 
>     I am confronted with a not so trivial segmentation task, for me at 
> least.  I have attached a slice from the 3D dataset.
> 
>     The structures I have to segment are tubular, and the section shows 
> a cross-section.  The upper tube is marked with a
> 
>     green dot, the lower tube is marked with two red dots (but it is one 
> tube, just for clarity).
> 
>  
> 
>     Now, I am able to get a nice initial segmentation using Curvature 
> Anisotropic Diffusion and Connected Thresholding.
> 
>     Then I experiment with the level-set algorithms using SNAP, starting 
> with the initial segmentation. It doesnt matter
> 
>     if I use a region or edge based speed-image, the two tubes always 
> leak into each other at the points where the edge
> 
>     between them is not so clear.
> 
>  
> 
>     Does anyone have an idea to get the edges more expressive (I already 
> tried gradient+sigmoid) or stop the leaking?
> 
>     Could it be possible to use a sort of competitive snakes, where I 
> use a seperate snake for each of the two tubes and
> 
>     the let both snakes evolve simultaneously and prohibit them from 
> intersecting?
> 
>  
> 
> Thanks In Advance,
> 
>  
> 
> Toon.
> 
>    
> 
> 
> ------------------------------------------------------------------------
>