Biomedical Engineering Reference
In-Depth Information
mammography [1]. The segmentation of anatomic structures — the partition-
ing of the original set of image points into subsets corresponding to structures —
may at least make contributions in the following applications:
1. Preprocessing for multimodality image registration, labeling, and motion
tracking. These tasks require anatomic structures in the original image to
be reduced to a compact, analytic representation of their shapes.
2. Finding anatomic structures of interest or foci for diagnosis and treatment
planning. A typical example is segmentation of the heart, especially the left
ventricle (LV), from cardiac imagery. Segmentation of the left ventricle
is a prerequisite for retrieving diagnostic information such as the ejection-
fraction ratio, the ventricular volume ratio, and heart output, and for wall
motion analysis, which provides information on wall thickening, etc. [2].
Another example is tumor segmentation.
3. Providing quantification of outlined structures and 3D visualization of
relevant image data.
Over the previous two decades, a number of researchers have devoted them-
selves to the development of eligible segmentation methods. Before analyzing
the state of the art of these methods, we first discuss several important issues in
medical image segmentation.
1.1.1. Complexity
Medical image segmentation is never a trivial task. Many factors should be
considered in the segmentation process. For instance, an obvious challenge is that
objects may be arbitrarily complex in terms of shape.
At present, various imaging modalities are in widespread clinical use for
anatomical and physiological imaging. A major hurdle in the effective use of these
imaging techniques, however, is reliable acquisition of anatomical structures. Ide-
ally, the image obtained should be sufficiently clear and free of artifacts to facilitate
diagnosis of pathologies. However, this requirement is never realized in practice.
Some effects often occur during the final images, including significant signal loss,
noise artifacts, object occlusion, partial-volume effects, and nonuniformity of re-
gional intensities. As mentioned below, these problems are readily present in
images acquired by brain MRI and cardiac ultrasound, where the boundary de-
tection problem is further complicated by the presence of confusing anatomical
structures. Occluded objects are often contained in ultrasound datasets. Partial-
volume effects are common in medical images, particularly with CT and MRI data,
which derive from multiple tissue type location within a single pixel (voxel), and
lead to blurring of intensity across boundaries. To compensate for these artifacts,
there has been a recent growing interest in soft segmentation methods [3-5]. Soft
