[Insight-users] Curvature term in ShapeDetectionLevelSetImageFilter

[Nils Hanssen]

Hi Luis,

yes, I am using the proposed pipeline [gradient magnitude] --> [sigmoid].
The feature image really looks like a binary image with high values inside
the tube and low values on the border. The problem are the gaps in the
border of the tube that have high values and act like "bridges" to the
surrounding high valued areas -> bad.
I thought, the regularizing terms in the level sets would be strong and
flexible enough to prevent the leakage, but for now I didn't find the good
parameters. Of course, I increased the curvature weighting and decreased the
propagation weighting simultaneously, but with no luck.

I choose the "axis" of the tube as the initial level-set, since this axis is
known "a priori" in my application.


Regards,
Nils


> -----Original Message-----
> From: Luis Ibanez [mailto:luis . ibanez at kitware . com]
> Sent: Wednesday, July 09, 2003 5:51 PM
> To: Nils Hanssen
> Cc: insight-users at public . kitware . com
> Subject: Re: [Insight-users] Curvature term in
> ShapeDetectionLevelSetImageFilter
>
>
>
> Hi Nils,
>
> In order to prevents leaks the ratio between the
> propagation scaling and the curvature scaling
> has to be modified.
>
> You may want to try reducing the propagation scaling
> factor as you increase the curvature scaling.
>
> The drawback in doing this is that if your initial
> level set is far from the actuar borders of the
> tubular structure that you want to segment, a low
> propagation scaling may be insufficient for bringing
> the level set to the expected edge.
>
> Note however that this balance of the scalings can
> only reduce the leaks, but will not prevent them from
> appening. As long as there is an open gap in the
> feature image, the level set will leak, it is just
> a matter of how many iterations it will take for it
> to leak.
>
> You may want to rework the feature image in order
> to make darker the edges. This can be done by reducing
> the "beta" parameter of the Sigmoid, and by using
> a smaller number (in absolute value) for the "alpha"
> parameter of the sigmoid.  The feature image should
> look almost as a binary imaga of the expected segmentation.
> (With almost black in the edges and almost white in
> the internal region to be segmented).
> (I'm assuming that you are using a sigmoid filter for
> preparing the feature image to be passed to the
> ShapeDetection filter...   is that the case ? )
>
> You may want to experiment also with letting the filter
> run for less iterations.
>
> --
>
> Please let us know if you continue experiencing problems.
>
> In that case it may be desirable to figure out a better
> way to initialize the filter for performing your segmentation.
>
>
> Regards,
>
>
>     Luis
>
>
>
> ---------------------
> Nils Hanssen wrote:
> > Hi,
> >
> > I want to segment a tubular object with the ShapeDetectionLSIF. The
> > initial level-set is the "axis" of the tubular object.
> > To prevent the contour from leaking through gaps in the
> edge-map, I set
> > the curvature-weighting with values between 0.02 and 5.
> > The size of the gaps is approx. 3% of the perimeter of the
> contour, but
> > I was not able to stop the leaking.
> >
> > What do I have to tune to achieve a significant smoothing?
> >
> >
> > Regards,
> > Nils
> >
> >
> >
> >
>
>
>
>