Biomedical Engineering Reference
In-Depth Information
set space. The brain ventricle extraction demonstrated the potential of the method
in a higher-dimensional space.
Freedman et al. [65] presented an algorithm that uses learned models for
both the shape and appearance of objects to achieve segmentation, which involves
learning both types of information. The authors the application of the algorithm to
segmentation of the prostate, as well as adjacent radiation-sensitive organs (e.g.,
bladder and rectum) from 3D computed tomography (CT) imagery, for the purpose
of improving radiation therapy. The main innovation of this method over similar
approaches is that there is no need to compute a pixel-wise correspondence between
the model and the image; however, appearance information could be fully utilized.
6.
CONCLUSIONS
Segmentation of deformable objects from medical images is an important
and challenging problem. In the context of medical imagery, the main challenges
include the following: (1) the edge information of the objects is too weak, (2) many
objects have similar intensity profiles or appearance, and (3) a lot of the objects
have a similar shape. Existing algorithms that use both shape and appearance
models require a pixel-wise correspondence between the model and the image;
this correspondence problem can sometimes be difficult to solve efficiently.
One type of segmentation algorithm incorporates learned shape models for the
objects of interest using level set methods. One advantage with this method is that
there is no need to compute a pixel-wise correspondence between the model and
the image. The algorithm allows the shape to evolve until the optimal segmentation
is found by minimizing a region-based energy factor.
In this chapter, this type of segmentation method was introduced in detail and
practical applications were demonstrated. The applications not only proved that
the method is an efficient and accurate solution in medical image analysis, but new
attempts were also made to improve the original algorithm.
Although there are many advantages using the above methods within a level
set framework, the computational cost is still a large barrier. The level set method
is famous for speed due to heap sorting and signed distance transformation, but
there are still many problems for advanced application in level set frameworks.
For one thing, the Mumford-Shah model is fast without using the level set method,
but, using the level set method, the iterative update of the level set brings a huge
burden in terms of computation. From the segmentation results in [58], the reader
will find that the approach is way too slow. (Like Figure 10 in [58], small images
(e.g., 100
100) took more than 144 seconds to obtain the final segmentation.)
One important reason is solution of the PDEs involved in the frameworks.
Efforts have been made to speed up the computation, e.g., estimation of the PDE
[66], a multi-resolution approach [37], etc. However, it is limited to one aspect
of the problem. Convergence of the gradient decent method is also playing a key
×
