Biomedical Engineering Reference
In-Depth Information
ume sequences render a manual procedure impossible. Efficient processing of the
information contained in ultrasound image calls for automatic object segmentation
techniques.
When it comes to echocardiographic sequences, the inherent noise, the simi-
larity of the intensity distribution between object and background, and the complex
movements make it difficult to segment the cardiac valve accurately. We could
make full use of the valve prior, and segment it under the guidance of prior knowl-
edge to reduce the manual intervention. We can segment the valve automatically
guided by the following prior:
1. The valve moves in a relatively fixed region between neighboring images.
2. The valve has a relatively stationary shape at a certain position.
We have developed an algorithm based on a level set framework, and repre-
sented the prior as a speed field. The speed field drives the zero level converging
on the ideal contour and then the valve of the heart is segmented. Section 5.2 gives
an overview of some of the existing prior-based object segmentation methods. The
proposed algorithm is described and formulated in Section 5.3. The application
results are presented in Section 5.4, and conclusions follow in Section 5.5.
6.2. Related Work
There have been some applications of prior-based image segmentation under
the snake framework [43] over the past several years. One can freely incorporate a
prior into the function as an energy item. Cremers et al. [103] established a prior
Point Distribution Model (PDM) of a certain object, then calculated the contour's
posteriori Bayesian probability, and made it an energy item to control the evolving
process. Ivana et al. [104] modified the internal energy to preserve the thickness of
the cardiac valve leaflet. However, in the snake framework, the object is described
as a serial point, which makes it hard to deal with the changing topology. The goal
of valve segmentation is to study the mechanism of movement, which makes the
cardiac wall that joggles with the valve as important as the valve. The valve leaflet
can be divided into several isolated parts in certain slices. All of this means that
the topology of the contour changes when the valve moves. So it is too hard for a
snake model to be applied to this task. And, what is more, the snake model needs
an initial outline close to the contour, which makes an automatic segmentation
process impossible.
Level set-based segmentation embeds an initial curve as the zero level set of
a higher-dimensional surface, and evolves the surface so that the zero level set
converges on the boundary of the object to be segmented. Since it evolves in a
higher dimension, it can deal with the topology changing naturally. However,
it is difficult to incorporate prior knowledge into the evolution process to make
the segment result more robust and efficient. Leventon et al. [105] calculated the
