Biomedical Engineering Reference
In-Depth Information
An additional cavity or handle-filling step is always required after normal segmen-
tation in order to get the right topology.
Region-based approaches are well known for their robustness, but they can
be computationally expensive (e.g., optimization and parameter estimation of the
MRF). These techniques can be further classified into many categories (see [26] for
the details). It is worth noting that the authors of [26] considered prior knowledge-
based techniques to belong to the region-based segmentation class, as the one that
we propose below to segment the cardiac valve in echocardiographic sequences.
1.2.2. Boundary-based methods
In traditional boundary-based methods, local discontinuities are detected first
(called
edge detection
) and then connected to form complete boundaries (called
edge linking
). There has been a great number of studies on edge detection over
the last four decades. Several well-known numerical techniques for intensity edge
detection contain Sobel's, Robert's, Kirsh's, and Prewitt and Canny's operators.
Several surveys on early edge detection work can be found in [38,39]. The use
of edge detection techniques often results in broken edges due to noise. It is
generally recognized that boundary detection is therefore finished by edge-linking
algorithms.
Readers interested in further information about this are referred to
[40, 41].
The biggest problem of edge detection may be that most traditional gradient
operators are sensitive to noise and produce spurious edge elements that make
it difficult to construct a reasonable region boundary. In addition, since the two
stages of edge detection and edge linking are independent, some missing edge
segments in the former stage could never be brought back.
Over the last two decades, probably the most visible approaches brought
to maturation in terms of both methodology development and application were
boundary-finding strategies based on deformable models [1]. Such a model can
obtain an accurate boundary of objects by deforming the initial curve, which is
defined in advance. The deformation process is guided by minimizing with respect
to the initial curve a functional, whose local minimum is given by the boundary of
objects.
Deformable models, though stemming from the computer vision field, are
frequently used in medical image processing, and have been widely applied to
problems, including segmentation, tracking, and registration. For example, Gupta
et al. [2] used deformable model-based techniques to segment the ventricular
boundaries in cardiac MR images. Rifai et al. [42] employed a 3D deformable
model with variable topology to segment the skull in MRI volumes, where they
also tok into account the partial-volume effect to formulate the speed function for
the model. Davatzikos and Bryan used an active contour to obtain a mathematical
representation of the cortex. These are but a few examples. A more detailed
