Biomedical Engineering Reference
In-Depth Information
set an initial contour for the object, and it will be driven to converge to the true
boundary.
Most problems of segmentation inaccuracy can be overcome by human in-
teraction. Even a correct automatic segmentation could not be obtained without
relying on prior knowledge about images. Promising segmentation methods for
complex images are therefore user guided and thus semiautomatic. An elegant
semiautomatic should find an appropriate compromise between manual interac-
tion and performance. In recent years, several effective interactive segmentation
tools have been developed for delineating the boundaries of objects, which require
manual intervention and guidance and consist of fast and accurate refinement tech-
niques to assist the human operator. These tools include intelligent scissors [18],
live wire/lane [19], graph-cut [20], and grab-cut [21]. It would will be meaningful
to apply them to medical image segmentation.
1.2. Segmentation Methods
Many segmentation methods proposed for medical-image data are either direct
applications or extensions of approaches from computer vision. Several general
surveys on image segmentation can be found in the literature [22, 23]. Several sur-
veys have targeted segmentation of MR images in particular [3, 24-26]. According
to four time frames, Duncan and Ayache [1] critiqued the research efforts that have
been put forth in the area of medical image analysis, including segmentation, over
the past 20 years or so. Pham et al. [7] presented a critical appraisal of the current
status of semiautomatic and automatic methods for the segmentation of anatomical
medical images. In their work, the current segmentation approaches are subdivided
into eight categories: (a) thresholding approaches, (b) region-growing approaches,
(c) classifiers, (d) clustering approaches, (e) Markov random field (MRF) mod-
els, (f) artificial neural networks, (g) deformable models, and (h) atlas-guided
approaches. Suetens et al. [27] reviewed different classes of image segmenta-
tion methods in diagnostic radiology and nuclear medicine. Based on the level
of implemented model knowledge, they classified these methods into (a) manual
delineation, (b) low-level segmentation, and (c) model-based segmentation.
Most of these methods are based on two basic properties of the pixels in
relation to their local neighborhood: discontinuity and similarity. From this logic,
in this chapter we coarsely classify the methods of medical image segmentation
into three subcategories: region-based methods, boundary-based methods, and
their combination. We will present brief overview of other notable methods that
do not belong to these categories at the end of this section.
1.2.1. Region-based methods
Region-based methods are techniques that segment the image/volume into
regions/sub-volumes.
Several commonly used techniques belong to this class:
