Biomedical Engineering Reference
In-Depth Information
analysis of the use of deformable models in medical image processing can be
found in [13].
In contrast to the traditional low-level and model-free image segmentation
techniques, such as region growing and edge detection, they offer several advan-
tages for the segmentation of anatomical organs, including the following [13]:
1. Deformable models can be used to obtain closed and smooth contours,
even if the images are noisy.
2. Certain deformable models are topology preserving, while others allow
changes in the topology without introducing difficulties related to param-
eterization.
3. Some deformable models (e.g., using Fourier descriptors) allow obtaining
an analytic mathematical description of the segmented contour rather than
a set of points.
4. Deformable models can be used for 3D reconstruction of anatomical
organs.
Since the introduction of the snake model by Kass et al. [43], many researchers
have done much work on improving its performance. One of the most interesting
issues is extending its capture range and reducing its sensitivity to initial placement.
In this aspect, various methods (e.g., the balloon [44], distance transformation [45])
have been proposed. One approach worth mentioning here is Gradient Vector Flow
(GVF) [46], which combines almost all the merits of the previous snakes. More
importantly, it achieves a large capture range by using a regularization term. It
moves snakes into boundary cavities that were unattainable by conventional snakes.
In this chapter GVF is applied to refine the contour in tongue segmentation. Snakes
are referred to as parametric deformable models in some of the literature [1, 47].
Corresponding to parametric deformable models are geometric deformable
models, which were first proposed by Osher and Sethian [48]. The relationship
between parametric a deformable model and a geometric deformable model was
explored in [47, 49]. Geometric deformable models are in the form of curve evo-
lution within the level set framework, where contour deformation is treated as a
propagating wavefront that can be seen as the zero level set of an evolving func-
tion. Geometric deformable models have been brought into the medical arena by
a variety of researchers [50-52].
1.2.3. Hybrid methods
There has been a significant trend in recent years to fuse region information
with a boundary detector [53-55]. One typical example is the level set method
with regularizers [56]. Regularizers take in various forms, including shape and
