[Insight-users] an error when running ShapeDetection

[Luis Ibanez]

Hi Zhao,

The segmentations produced by evolution of LevelSet
are necessarily narrower than the actual anatomical
structures.

The reason is that the speed of the wavefront is
usually proportional to a gradient magnitude.

Good gradients are computed after smoothing the
original image, (typically with a gaussian filter).
Hence, the resulting gradient magnitude has a certain
width which is proportional to the sigma of the
gaussian used to smooth the image.

Since the speed of the wave front is proportional
to the magnitude of the gradient, it is certain that
the wavefront will not cross the contours. However,
the fact that the speed is progressively reduced as
the wavefront approaches the contour make that the
wavefront will never get to the actual location of
the contour.

"The closer you get, the slower you advance..."
so the front will never get there.

A bit with the flavor of Zenon's paradox (The Dicothomy):
http://abel.math.harvard.edu/~knill/math21a2000/zeno/


--

Possible alternatives are:

1) Reduce the smoothing of the derivative filter...
   but this is restricted by the noise in your image...
   if you don't smooth enough, the wave front will
   probably get stock in small secondary details and
   image texture.

2) Take the segmented image and apply a Morphological
   Dilation with a structuring element of radius equal
   to the sigma used in the gradient smoothing filter.
   This assumes that the wavefront will virtually stop
   at a certain distance of the actual contour, so you
   are entitled to extend the object a certain distance.


Please let us know if you encounter other difficulties,

Thanks


 Luis


-----------------------

zhao wrote:

> Hi Luis,
>  	
> 	OK now. Thanks a lot.
> 	But it seems that the extracted shape of ventricles is apparently narrower than the visual one in the original image, Which is true even after I greatly increased the "t" parameter. 
> 	Other segmentation results derived from different seed points also seem mismatch with original anatomy structures.
> 
> Zhao
>    
>