Biomedical Engineering Reference
In-Depth Information
identifying cellular structures and cell motion. This wide range of applications
essentially covers both the 2D and 3D image domain for signifying both volume
and temporal data. Motion tracking using 3D temporal data has also been widely
studied for cardiac, pulmonary, and arterial motion in 4D MR, CT, etc., and even
to the level of cellular motion from molecular imaging devices. At the present
time, the deformable model-based segmentation algorithm has become a vital part
of the most advanced image processing toolbox associated with medical imaging
devices.
Most clinical applications presently use manual segmentation of the region of
interest, that is, a domain expert goes through each of the image slices over the en-
tire volume, or temporal data, and manually identifies and delineates the region of
interest by using a mouse-guided framework. This has several disadvantages: the
manual segmentation process is extremely tedious and time consuming. Further-
more, the image segmentation is nonreproducible and prone to operator bias. Thus,
computer-assisted methodologies with minimal user intervention need to replace
the manual segmentation process to obtain accurately reproducible segmentation.
This chapter has attempted to focus on the development of the active contour
model to meet the various requirements in medical image segmentation and analy-
sis. Segmentation of medical images is required for accurate and reproducible
analysis of data for a huge range of applications, including diagnosis, postop-
erative study, and interactive surgical procedures. Manual segmentation is the
most common technique used by physicians to process data. However, with the
amount of data exploding and due to the nonreproducibility of the results, there
is an inherent need for automated or semi-automated computerized algorithms
that can generate segmentation results accurately and reproducibly. Segmentation
processes that use low-level image processing have not been sufficient to segment
complex structures from images and provide an accurate continuous boundary due
to their dependence on local statistics, which in turn are corrupted by noise and
other artifacts. The deformable model has been found to be quite efficient in this
context, since it uses physics-based constraints along with local image statistics
in a very natural way. The initial design of the active contour itself generated a
lot of interest. With the complexity of the task increasing, the requirements are
becoming more demanding, and subsequent improvements have followed. Efforts
are being made to make the deformation less sensitive to initialization [13, 26].
Dependence on the gradient force alone for growth and termination of the active
contour forces the snake to fail in images with weak and fuzzy boundaries, since
the edge functional is not well defined in those regions. Regional information,
like homogeneity and contrast, improved the active contour model. The snake
evolution using regional information and local statistics, like gradient, captured
the object effectively as the propagation was controlled by the regional force rather
than a blind force. A-priori information even made the snake bidirectional, thus
helping it to prevent leakage. Performance was thus enhanced with incorporation
of this form of energy. Another major advancement of the active contour model
