Biomedical Engineering Reference
In-Depth Information
Based on the model described above, we further present a unified approach for
semiautomatic segmenting and tracking of the high-resolution color anatomical
Chinese Visible Human (CVH) data. A two-step scheme is proposed, starting with
segmenting the first image slice. The user can initialize the first slice by clicking
on the area of interest (AOI). The contours in the subsequent image slices are then
automatically tracked. The underlying relationship of these two steps relies on
the proposed variational framework for the speed function. Segmentation results
are shown to be promising, while the tracking procedure demands very little user
intervention.
The rest of this chapter is organized as follow. Section 2 discusses the related
work. Section 3 covers the mathematical foundations of level set methods. We will
discuss the fundamental formulation first. We will then look into why the curvature
term is combined in the speed function, and how to adapt the speed function for
image segmentation. The front propagation algorithm of the
narrow band
method
and the numerical scheme are also mentioned in this section. In Section 4 we
present the work on incorporating the speed function with the relaxation factor
and the image content items to handle tagged cardiac MRI images. In Section 5
we present the semiautomatic segmentation and tracking of serial medical images.
A two-step scheme is proposed to reduce the user intervention. A new scheme to
construct the speed function for tracking is demonstrated with the high-resolution
color anatomical images. Finally, we offer our conclusions in Section 6.
2.
RELATED WORK
Medical image segmentation plays an important role in computer-aided diag-
nosis. Because of the complexity of organs and the structures and the limitation
of different imaging modalities, medical images always show artifacts, includ-
ing noises and weak boundaries. Such artifacts make medical image segmenta-
tion a difficult task. When applying classical segmentation techniques (such as
thresholding, edge detection, and region growing), these artifacts may cause the
boundaries to become indistinguishable and discontinuous. Figures 1a and 1b
show typical tagged cardiac MRI and their corresponding edge image obtained
using the Roberts operator. In Figure 1a, because of the blood flow, there is an
artifact in the image of the left ventricle. In Figures 1a and 1b, the graylevel of
boundaries varies from place to place, and these boundaries are not apparent, and
hence cannot be detected by the Roberts operator. Examining Figures 1c and 1d,
the segmentation results obtained using the Roberts operator, we can see that there
are discontinuities on the boundaries, and they cannot be correctly segmented. As
a result, these model-free techniques either fail completely or require a significant
amount of expert intervention for such a segmentation problem.
To address these difficulties,
deformable models
have been proposed and are
now widely used in medical image segmentation [5]. Deformable models are
