Biomedical Engineering Reference
In-Depth Information
accurately segment the objects of interest. Image segmentation is the process of
delineating an image into several “homogeneous” regions based on the similarity of
pixel attributes. If the segmentation is performed independently by the computer,
the processing is referred to as
unsupervised
. This is in contrast to
supervised
image segmentation, which requires human input and intervention. Depending on
whether or not prior knowledge has been used in the image segmentation, it can
be classified as either “high-level” or “low-level” segmentation. Low-level image
segmentation relies upon the pixel attributes of the image without consideration of
prior knowledge. For many practical applications, however, it is often necessary
to incorporate additional information into the analysis.
New segmentation methods have emerged that guide the partitioning process
by utilizing cues based on shape, appearance, and/or contextual models. In early
low-level segmentation pixels are clustered based on their similarity in spatial
and feature space and the resulting subregions are then assigned object labels. In
contrast, high-level approaches attempt to detect and extract objects from images
using prior knowledge and establish models that allow the segmentation method
to adapt to the object of interest. Deformable models belong to this family of
high-level image segmentation approaches.
There are two general types of deformable models described in the literature:
parametric deformable models [4] and geodesic or level set-based deformable
models [5, 6]. Parametric deformable models have gained significant attention
throughout the image-processing community since its first introduction by Kass,
Witkin, and Terzopoulus [4]. Snakes are curves that are defined within the image
domain and move under the influence of internal forces within the curve and ex-
ternal forces derived from the image data. All deformable model properties and
behaviors are specified through a function called the
energy function
by analogy
with physical systems. A partial differential equation controlling the deformable
model causes it to evolve so as to reduce its energy, and the local minima of this
energy then correspond to desired image properties. Parametric deformable mod-
els have been used in a range of applications, including edge detection [4], object
recognition [7, 8], shape modeling [8, 9], and motion tracking [7, 10], to mention
only a few. In an almost parallel effort, a variety of deformable models based on
utilizing the image gradients as external forces have been proposed. Examples
include the traditional deformable model [11, 12], the balloon-deformable model
[13], the pressure forces model [14], and the more recently reported gradient vec-
tor flow (GVF) model [15, 16]. The GVF deformable model often outperforms
other gradient-based models because it is insensitive to initialization values and
can move into boundary concavities. It also has a much larger capture region than
earlier approaches. However, the GVF deformable model was designed for binary
or graylevel images, and it is not straightforward to adapt this approach to segment
imaged pathology specimens. Simply transforming color images into graylevel
images suffers from the fact that this process can often serve to eliminate potentially
useful chromatic attributes, which may contain extremely valuable informational
