[Insight-users] REPOST - itkGeodesicActiveContourShapePriorLevelSetImageFilter/PCAShapeSignedDistandeFunction

Lydia Ng lydiang at gmail.com
Mon Nov 22 19:27:40 EST 2004


Hi Karl,

>From my experience with both toy and some real corpus callosum
datasets, I found that the ITK implementation (and I think Leventon's
implementation also) is sensitive to the initial location of the
"best-fit" model.

The initial location should be such that at least there is an overlap
between initial "best-fit" shape and the initial level set. This is
because the ShapePriorMAPCostFunction is only computed within a band
near the zero-level set. If you look at Leventon's paper (CVPR2000)
you can see this is the case for the 9 examples he shows in Figure 9
to Figure 12.

Also it is my opinion that to get good result you should place you
initial "best-fit" shape to cover as much as the object you want to
segment as possible. I believe this is what Leventon has done in the
examples in Figure 10 - 12.
For example, in the the top-left image of Figure 11. I believe there
is no way for the algorithm to get such as good estimate of the
"best-fit" shape based on the small circle initial level set alone -
the "best-fit" must have been initialize pretty close to begin with.
But that is only my guess from the paper - if anyone has access to
Leventon's implementation, I would be very interested to hear about
their experience.

By the way - if you have not already seen there is an example in
Segmentation/GeodesicActiveContourShapePriorLevelSetImageFilter.cxx
using some synthetic corpus callosum data.

Hope this helps,
Lydia

On Mon, 22 Nov 2004 10:39:49 +0100, Karl Fritscher
<karl.fritscher at umit.at> wrote:
> Hi ITK-users,
> 
> I m using the itkGeodesicActiveContourShapePriorLevelSetImageFilter
> and have one question to setting the initial parameters. Setting the
> initial paramters for the  transformation in the
> PCAShapeSignedDistandeFunction would mean that I do have to know the
> exact position of the object to be segmented,right?! As far as I
> remember Leventon s implementation , the model has been aligned to the
> evolving contour during the segmentation process. Is this also the case
> in the ITK implementation?!(Or what else is the optimizer in the filter
> doing?)
> Would this mean that i only need rough estimates for the initial
> parameters?
> 
> Many thanks in advance
> 
> Karl
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
>


More information about the Insight-users mailing list