Biomedical Engineering Reference
In-Depth Information
thresholding-, clustering-, and region-growing-based segmentation, and MRF
models.
Thresholding is the most intuitive approach to segmentation [28]. It maps
and clusters pixels in a feature space called a histogram. Thresholds are chosen at
valleys between pixel clusters so that each pair represents a region of similar pixels
in the image. Interactive selection of thresholds for the intracranial region has been
reported [29]. This works well if the target object has distinct and homogeneous
pixel values. On the other hand, spatial information is lost in the transformation,
which may produce disjoint regions [30].
The difference between regions can be not only characterized by graylevel
values, but also represented in other derived statistical parameters. Therefore, we
can integrate various features — including intensity, and texture — into multidi-
mensional cluster vectors in order to distinguish different regions. A commonly
used example of clustering in medical image segmentation is the fuzzy
c
-means
(FCM) algorithm [5], where fuzzy memberships are constructed to segment MR
images that have been corrupted by intensity inhomogeneities. Problems similar to
histogram thresholding are anticipated, and in practice only well-defined regions
can be robustly identified.
Region growing [31, 32] is a classical segmentation technique extensively used
in computer vision. Starting with selection of a seed region, the region-growing
approach tends to group all similar neighbors according to a predefined homo-
geneity measure. Some early work on seed growing has been described by Cline
et al. [33], who used seed growing to extract the brain surface. Region growing
is seldom used alone, but is followed by a split-and-merge operation during the
entire process of image segmentation. In this scenario, a threshold is required to
decide whether to merge or split. The main disadvantage of region growing lies
in selection of the seed. It can be chosen empirically by the user, which may be
difficult if he there is no clear idea of the growth behavior of the region.
The basic idea of the MRF in computer vision was put forward by Geman and
Geman [34] in 1984, when they applied statistical mechanics to image processing.
The MRF is a conditional probability model multiplied by a priori probability per
se. In most cases, the conditional probability is in Gaussian form, representing
the possibility of the observed data given its actual value. The prior probability
models the spatial constraint in the neighborhood. Over the past twenty years,
extensive efforts have focused on application of an MRF model to various fields
[35]. Introduction of the MRF model to medical image segmentation was begun
by Wells et al. [8], who used a non-parametric Pazen window to model brain
images. One of the most significant contributions of their work is the correction
of RF inhomogeneity. Since then, this trend has been most significant [12, 36, 37].
As the MRF is a pixel classification method, the optimization and parameter es-
timation algorithms are often computationally expensive. Another drawback of
the MRF method is its inability to retain topological information. For example,
in cortical segmentation the MRF does not preserve its known spherical topology.
