Geometric Model Reconstruction - Computational Hemodynamics: Theory, Modelling and Applications - page 46

3.3.1 Segmentation Approaches

A region within an image can be defined by its pixel properties (e.g. greyscale in-
tensity), boundary (edge) or its interior. Therefore segmentation approaches can be
broadly categorised into the following:

• Pixel based (Thresholding) : each pixel is labelled based on its grayscale values
that represent intensity from the scans.
• Edge based : detects edge pixels to form a boundary containing the region of
interest
• Region based : considers pixel greyscale levels from neighbouring pixels by in-
cluding similar neighboring pixels (region growing).

Pixel based methods are the simplest and easiest approach to implement, however
they lack contextual information and fail in scans with high inhomogeneity through a
single region. Edge based methods are the next simplest approach and are efficient on
scans of anatomical structures that have clearly defined boundaries such as the artery.
A common problem however is that noise or occlusions can cause false or missed edge
detection. Region based methods are the most complete but complex methods since
regions of interest includes more pixel categorization than edges. Furthermore region
growing techniques are useful in noisy images where edges are difficult to detect.

3.3.2 Threshold Segmentation

Image segmentation is defined as the partitioning of an image area or volume into
non-overlapping, connected regions which are homogeneous with respect to some
characteristic such as intensity (Pham et al. 2000). The simplest form of segmenta-
tion is the selection of pixels in 2D (or voxels in 3D) based on criteria typically its
greyscale level, referred to as thresholding. This uses either global or local informa-
tion to select only those pixels within a greyscale range (threshold) and a binary
function is applied as,

( ) ( )

1 if min ( ) max ( )
0

fx ≤≤ θ fx

( )

x x

gx =

(3.1)

where θ is the selected greyscale value or a range of values and f(x) is the greyscale
values of the images. The output is a binary image where each pixel is coloured as
black or white, depending on a pixel's label of 1 or 0. It is most effective for im-
ages containing different structures that have uniform but contrasting intensities to
other quantifiable features or structures. The segmentation problem becomes one of
selecting the proper value for the threshold, θ. In many cases θ is chosen manually
by the user, by trying a range of values and observing the range that best identifies
the anatomical structure of interest.