Biomedical Engineering Reference
In-Depth Information
1.
INTRODUCTION
Breast cancer is one of the leading causes of death in women during recent
years (see
Cancer Facts and Figures 2004
, published by the American Cancer
Society [1]). The role of ultrasound in breast cancer detection has become more
prevalent as it offers several advantages for cancer detection, including its low cost
and non-ionizing nature (see Ganott et al. [2] and Crystal et al. [3]). For this reason,
considerable progress has been made in early cancer detection and diagnosis (see
Bassett et al. [4], Jackson [5], and Stavros et al. [6]). Recently, it has begun to
yield low false negative rates. Thus, due to all of the above reasons, ultrasound
has become and an indispensable modality [7-14].
Today's breast CAD systems require segmentation of breast tumors to analyze
the shapes of tumors, and then to classify these types of breast cancer diseases into
benign and malignant types. Thus, this introduces the prerequisite of accurate
breast boundary estimation in 2D breast ultrasound images and 3D breast ultra-
sound slices. However, boundary estimation of breast tumors is a challenging task
because there is no set pattern of the progression of tumors in the spatiotemporal
domain. The shape and structure challenge due to the anatomic nature of breast
diseases is just one part of the equation; the other component is the physics in-
volved in imaging these complex structures, as it produces several kinds of artifacts
and speckle noise. The anatomic nature of tumors causes blurry boundaries in 2D
ultrasound images, and there is considerable signal dropout. This puts an extra
burden on CAD (see Rohling et al. [15]).
This chapter adapts a methodology based on geometric deformable models
such as level sets that has the ability to extract the topology of the shape of breast
tumors. It has been recently shown that medical shape extraction in medical
imaging has been dominated by the level set framework (see Suri et al. [16-18]).
Recently, many implementations on computer vision and medical imaging have
been grounded on this basic concept [16, 19-26].
In the level set paradigm, tracking of the evolution of contours and surfaces
is solved using numerical methods. The efficacy of this scheme has been demon-
strated with numerical experiments on some synthesized images and some low-
contrast medical images (see Malladi [20]). A level set can be used for image
segmentation by using such image-based features as mean intensity, gradient, and
edges in the governing differential equation. In a typical approach, a contour is
initialized by a user and then evolved until it fits the topology of an anatomical
structure in the image (see Sethian [19]).
In this chapter we present several studies on using the level set method in
breast ultrasound. In order to increase the efficiency of level set-based segmen-
tation within the level set framework, some image preprocessing techniques are
adopted prior to application of the geometric level set. For performance evaluation
of the proposed studies, we not only show the segmentation results with the level
set method but also apply these results to follow-up analysis of breast tumors,
