Information Technology Reference
In-Depth Information
but difficult to achieve, as the complex cognitive abilities can hardly be trans-
ferred to computer programs. That is why there is no one standard segmentation
technique that can produce satisfactory results for all image applications. The
definition of the goal of segmentation varies according to the goal of the study and
type of data. Different assumptions about the nature of the analyzed images lead to
the use of different techniques. These techniques can be classified into two main
categories: region segmentation techniques that look for the regions satisfying a
given homogeneity criteria, and edge-based techniques that look for edges be-
tween regions with different characteristics [9]. Thresholding is a region-based
method in which a threshold is selected and an image is divided into groups of
voxels with values less than the threshold and groups of voxels with values greater
than or equal to the threshold.
3.1 Thresholding
Since segmentation requires classification of voxels, it is often treated as a pattern-
recognition problem and addressed with related techniques. The most intuitive ap-
proach for segmentation is global thresholding, when only one threshold based on
the image histogram is selected for the entire image. If the threshold depends on
local properties of some image regions, it is called local. If local thresholds are se-
lected independently for each voxel (or groups of voxels), thresholding is called
dynamic and adaptive.
For images that have biomodal histogram (i.e. grey levels grouped into two
dominant sets, object and background), the object can be extracted from the back-
ground by a simple operation that compares image values with the threshold value
τ
. Suppose an image
G
(
x
,
y
,
z
) with a histogram shown on the Fig. 2. The thresh-
old image
L
(
x
,
y
,
z
) is defined as
⎩
⎨
⎧
1
G
(
x
,
y
,
z
)
>
τ
L
(
x
,
y
,
z
)
=
.
(10)
0
G
(
x
,
y
,
z
)
≤
τ
The result of thresholding is a binary image, where voxels with threshold value 1
correspond to objects, while voxels with value 0 correspond to the background.
There are a number of selection methods for threshold
based on classification
model that minimizes the probability of an error. With the semiautomated version,
an expert (operator) selects two voxels - one inside an object and one from the
background. By comparing the distribution of voxel intensities in the circular re-
gions around the selected voxels, the threshold is calculated automatically. It cor-
responds to the least number of misclassified voxels between two distributions.
The result of the thresholding operation is displayed as a contour map and super-
imposed on the original image. If needed, the operator can manually modify any
part of the border.
If an image contains just two types of regions, objects with uniform intensity
values and a contrasting background, in most cases good segmentation is obtained
when the background area and the objects are minimally sensitive to small varia-
tions of the selected threshold level. In this case global thresholding can be
τ
