[Insight-users] On the convergence speed of level set methods

[Joshua Cates]

Hi,

The best way to deal with this issue is to do a multi-scale segmentation
with level-sets, or initialize your level-set segmentation with a prior
generated by some faster method (confidence connected, for example). A
multi-scale approach uses downsampled volumes at progressively higher
resolutions to produce successively larger initializations.  By the time
you reach the full resolution, your initialization will be very close to
the final solution, requiring much few iterations.

There are a few slides on this in the tutorial:

http://www.itk.org/CourseWare/Training/SegmentationMethodsOverview.pdf

In general, level-sets are most effective when given an initialization
that is reasonably close to the solution.  Different methods will be more
or less sensitive to initial conditions, but you will find some methods to
be useless if initialized from a "seed point" surface.  This is because
the solver might converge on many reasonable solutions between your seed
point and the solution that you are after.  

Another approach is to run the code in parallel.  Both the narrow band 
level-set solver and the sparse field level-set solver can take advantage 
of multiple processors.

Josh.


On Wed, 7 Apr 2004, Kai Li wrote:

> Hi,
>    I tried the level set method (GeodesicActiveContour) in ITK. I found
> that when the front grows larger and larger, the time for each iteration
> becomes slower and slower. It finally takes a long time for the whole
> process to converge. This is especially the case for segmenting
> shapes like the white matter. I'm investigating methods for improving this
> problem. The question I'd like to ask here is whether ITK already has any
> effective mechanism for this problem.
> 
> Thanks,
> Kai
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
>