Biomedical Engineering Reference
In-Depth Information
object into the segmenting process makes the ambiguous, noise-stained, or occluded
contour clear, and the final result becomes more robust, accurate, and efficient.
The deformable models to be described below make it easy to incorporate prior
knowledge to segment certain objects, where prior knowledge may be incorpo-
rated into models in the form of initial conditions, data constraints, or constraints
on the model shape parameters [13].
Initially, deformable templates were used to embody a priori knowledge of the
expected shape and shape variation of the structures. The idea can be traced back
to the early work on spring-loaded templates by Fischler and Elshlager [14]. An
example in medical image analysis is the work of Lipson et al. [15], who extracted
vertebral contours using a deformable ellipsoidal template. Since then, more and
more researchers have incorporated shape priors into their models to segment
medical images. Staib and Duncan [16] applied a probabilistic deformable model
on 2D echocardiograms and MR images to extract the LV of the heart and the corpus
callosum of the brain, respectively. Probability distributions on the parameters of
the representation bias the model toward a particular overall shape while allowing
for deformations. Later, Staib and Duncan extended the model to 3D [17].
1.1.3. Interaction
A wide variety of approaches have been proposed for medical image seg-
mentation. These approaches can be considered as three broad classes: manual,
semiautomatic, and automatic. Currently, in most clinical segmentation, a skilled
operator, using a computer mouse or trackball, manually traces the structure of in-
terest on each slice of an image volume. Performing this segmentation manually is
extremely labor intensive and time consuming, which is not even feasible for some
applications with the increasing size of datasets. There are additional drawbacks
in terms of achieving reproducible results due to operator bias and fatigue.
On the other hand, automatic segmentation of medical images can relieve
clinicians from the laborious and tedious aspects of their work while increasing
the consistency and reproducibility of segmentation. Although efficient, an au-
tomatic segmentation still aims at excellent performance. From this viewpoint,
semiautomatic segmentation provides a good tradeoff between the precision of
manual segmentation and the efficiency of automatic segmentation. Semiauto-
matic methods require user interaction to set algorithm parameters, to perform
initial segmentation, or to select critical features. Many existing segmentation
methods fall into this class. For example, when segmenting bony structures in
CT images, a commonly used method is histogram thresholding, where one or
more thresholds are selected from the histogram to group the pixels into several
different clusters. In region-growing algorithms, seeds are generally chosen by the
user, from which the final regions are formed. (Note that most of the deformable
models used in this chapter are semiautomatic methods.) The user is required to
