Geometric Model Reconstruction - Computational Hemodynamics: Theory, Modelling and Applications - page 45

3.3 Image Segmentation

Segmentation is the process of extracting one or more contiguous region of interest
representing individual anatomical objects based on a discontinuity or a similar-
ity criterion. More precisely, it is the process of assigning a label to every pixel
based on a criterion in an image so that pixels with the same label and that are
connected forming a contiguous region can be extracted, identified or categorised.
Furthermore neighbouring pixels which do not belong to the anatomical structure
will exhibit a pixel value outside of the criterion.
Segmentation can be manually performed by selecting individual pixels on the
cross-sectional slices or automated. In manual segmentation the user selects the
region of interest in every cross-sectional image in a set of scanned images which
can have up to 1000 images. Therefore this procedure is time-consuming and also
lends itself to inter-observer and intra-observer variability. Fully automated or semi-
automated segmentation for monochrome images are generally based on disconti-
nuities or similarities within the image based on greyscale values (referred to as
intensity). For example, discontinuities at edges in an image can be identified based
on abrupt changes in the intensity (greyscale level). Similarity in the intensity is
used to extract a region which exhibits similar properties according to a set of pre-
defined criteria.
Segmentation methods have been explored for many years producing a large
number of algorithms dependent on the specific application, imaging modality, and
other factors. For example segmentation of carotid arteries involves handling of
outliers, feature point detection, and additional user interaction. However segmen-
tation remains a challenging problem to overcome the increasing number of ana-
tomical structures of interest, large variations in the properties within the images,
and imaging artefacts such as noise, partial volume effects, and motion-blur. There-
fore there is no single algorithm that can produce sufficient results for all types of
medical images.
A segmentation issue related with MR or CT imaging is intensity inhomogene-
ity artifact (Condon et al. 1987; Simmons et al. 1994; Sled and Pike 1998), which
can cause problems with algorithms that assume a constant intensity value for a
tissue class. Performing a pre-filtering operation can remove the inhomogeneity
by assuming mean tissue intensity for each tissue class is spatially varying within
a certain range.
This section provides the reader with an introduction to the field of image seg-
mentation and a description of some of algorithms among the many that exist in the
literature. For further readings, there are general reviews on the segmentation of
MR images (Bezdek et al. 1993; Clarke et al. 1995; Liang 1993; Peters et al. 1993),
comparisons of different segmentation methods for MR images (Clarke et al. 1993;
Hall et al. 1992; Vaidyanathan et al. 1995), segmentation of CT images (Sivewright
and Elliott 1994), while CT segmentation applications of the abdominal aortic an-
eurysms (Juhan et al. 1997), carotid artery (Zhu et al. 2013) and segmentation of the
heart (Ecabert et al. 2008), and cerebral artery (Manniesing et al. 2008).