[Insight-users] ConfidenceConnectendImageFilter

Miller, James V (Research) millerjv at crd . ge . com
Tue, 6 May 2003 09:46:57 -0400


The number of iterations to use is largely based on the nature
of the data.  So I'll give a little background on the algorithm
to aid the discussion.

The algorithm starts with a seed point and calculated an sample
mean and sample standard deviation in a neighborhood of the seed point.
The algorithm the grows a region around the seed point until it can
no longer find adjacent pixels that fit the original sample statistics.

Given the size of the initial neighborhood, the sample statistics
may have a large amount of uncertainty.  So the algorithm recalculates
the sample statistics using the first segmentation.  The it starts
over at the original seed point and segments again.  The process
continues for each iteration.

How many iterations should you run?  If the statistics in the region
you are trying to segment are uniform across the region, then my gut
feeling is to run 3 iterations.  The first iteration invariably is 
biased by the uncertainty in the sample estimates from the initial 
small neighborhood.  The second iteration uses a larger sample size
so the estimates are better but are biased by the first segmentation
which is biased by the initial neighborhood.  By the third iteration, 
most of these effects are minimized.

This rule of thumb however is not always appropriate.  If you are
trying to segment something with a subtle boundary, I would reduce
the number of iterations (perhaps to 1).  If the data across the region
has statistics that vary across the region (but are distinct from the 
background), then a larger number of iterations will be needed. In these
cases, I have used as many as 10 iterations.



> -----Original Message-----
> From: salah [mailto:salah@gris.uni-tuebingen.de]
> Sent: Tuesday, May 06, 2003 9:21 AM
> To: ITK Users
> Subject: [Insight-users] ConfidenceConnectendImageFilter
> 
> 
> 
> Hello,
> 
> what could be a good choice for the number of iterations 
> to be used with the ConfidenceConnectendImageFilter.
> 
> Thanks,
> Zein
> 
> ->8<------------->8<------------->8<------------->8<----------
> --->8<------------->8<-
> Zein I. Salah 
> University of Tubingen, WSI-GRIS, Sand 14, 72076 Tubingen 
> Email: salah@gris.uni-tuebingen.de
> Tel.: (07071) 29 75465 (GRIS),           Fax: (07071) 29 54 66
> 
> _______________________________________________
> Insight-users mailing list
> Insight-users@public.kitware.com
> http://public.kitware.com/mailman/listinfo/insight-users
>