[Insight-users] a method question about GeodesicActiveContourShapePriorLevelSetImageFilter

[Haiyong Xu]

Hi there,

I have a question about the segmentation method that using shape prior
and levelset. The segmentation problem I encountered is that there is
a small round protrusion (10x10x10 mm) on a flat surface. The
protrusion and the base (the flat surface) have almost the same CT
intensity values (soft tissue), but the area around the protrusion has
very different CT intensity (air). I want to segment ******the
protrusion part only******. Since there is no intensity difference
between the protrusion and the base tissue, the levelset method itself
cannot do it. Therefore, I build a shape prior model to simulate the
protrusion. As in
itk::GeodesicActiveContourShapePriorLevelSetImageFilter, this shape
prior model is a signed distance map, which is defined by a mean and
variation mode of training samples. In order to construct a signed
distance map, I artificially introduce a base surface for the
protrusion part on each training sample so that the zero levelset in
the signed distance map is a closed contour. Now, if I place the zero
levelset of shape prior model on the image which is going to be
segmented, the protrusion part of the shape prior model would nicely
match the image feature, and the base part of the shape prior model
would be emerged in the soft tissue. Using
GeodesicActiveContourShapePriorLevelSetImageFilter, I expected the
image feature could stop the contour expanding for the protrusion
part, and the shape prior model could stop the contour expanding for
the base part.

I use the example of
GeodesicActiveContourShapePriorLevelSetImageFilter in ITK and I got
some results which show that the method works, in a certain degree.
Without shape prior model (I set the parameter of shapePriorScaling
equal to zero), the contour expands very far to the surround soft
tissue. With the shape prior model, the final contour does not expand
very far in the soft tissue. The problem is that although the
protrusion part of contour could be stopped exactly on the boundary,
which is very nice, the base part is different. The shape prior model
does provide a support, but can not provide enough force to stop the
contour exactly on the artificial base surface which was introduced in
the training samples. I tried to adjust the weight of shape prior
model in order to balance the forces of propagationScaling, which
comes from image features, and the shapePriorScaling, which comes from
the shape prior model. It did NOT provide very good results. My
question is that is there a better way to use the
GeodesicActiveContourShapePriorLevelSetImageFilter to solve my problem
or a different approach to do it. I appreciate any suggestions.
Thanks.

Best,
Haiyong Xu