Segmentation of B
Using
Abdel–Razzak Natsheh, Prasad VS Po
Manchester
*
Traffo
a.natsheh@mmu.ac.uk n.anani
Abstract—The diagnosis of sinus diseases (e.g.
requires the segmentation of bone structure
sinus areas in CT (computer tomography)
uniformity of bone tissue, ranging from dens
textured spongy bone, the irregular shapes
bones, and the small inter-bone spaces
resolution of a CT image coupled with the
vessels and the inherent blurring of CT im
segmentation of a bone a challenging task. In
technique is presented which uses hierarchi
maps to segment the structure of bone surr
regions. The algorithm has been successfull
tested with various CT images that have
disease.
I. INTRODUCTION
Diagnosis of a sinus condition is a proces
use of multiple sources of information an
laboratory and visual tests, computerized t
scan images, and the expertise of a clinician
sinus condition diagnoses is the quality and
images that a clinician uses to make decision
differ significantly in terms of quality and a
The diagnoses can be further complicated by
patients’ anatomy, the extent of diseased s
bone structure. Discussions with ENT
radiologists highlighted the need for an aut
based analysis and diagnosis tool that can be
an aid to experts, but also for use in the train
medical students. Such a tool, the Sinus D
henceforth referred to as the (SDS) has been
techniques based on classical image processi
procedures in conjunction with Artificial
(ANNs) tools. Previous work reported the po
using ANNs in enhancing early diagnosis [1
important stages in developing the propos
image segmentation stage in extracting the b
the CT images [1].
Medical image segmentation, however, h
demanding task. This is true, in particular, fo
of bones within the sinus areas in CT i
general, the segmentation of bones in CT im
a relatively straightforward task, segmen
surrounding sinus areas is tricky becaus
datasets contain irregularly shaped bones wit
distances relative to the resolution of CT
known fact is that bone tissue cannot
Bone Structure in Sinus
g Self-Organizing Maps
nnapalli, Nader Anani, Dalil Benchebra, Atef El-
r Metropolitan University, Manchester M1 5GD, UK
ord General Hospital, Manchester M41 5SL, UK
i@mmu.ac.uk , p.ponnapalli@mmu.ac.uk , atefelk
malignant disease)
e surrounding the
images. The non-
se cortical bone to
of closely packed
compared to the
presence of blood
maging, render the
this paper, a novel
ical self-organizing
rounding the sinus
ly applied to, and
different types of
ss that involves the
nd skills such as
tomography (CT)-
n. A key feature of
complexity of the
ns. The images can
angles of imaging.
y the irregularity of
sinus area and the
consultants and
tomated computer-
e used, not only as
ning of doctors and
Diagnostic System,
n developed using
ing algorithms and
Neural Networks
otential benefits of
]. One of the most
sed system is the
one structure from
has proved to be a
or the segmentation
mages. While, in
mages is considered
ntation of bones
se the volumetric
th small inter-bone
imaging. A well
be characterized
uniformly: the outer layer of th
denser than the spongy bone n
eyes. Thus, under CT imaging
smooth while spongy bone app
blood vessels resemble the back
surface of bone images. I
characteristics translate into
segmentation techniques (see f
shell (between the eye and sin
bone boundaries due to the
imaging; (iii) textured areas co
alternating between bone-like
(iv) the narrow inter-bone regio
Figure 1. Weak e
While it is possible for me
images using thresholding or
ANALYZE package or simil
intensive considering the time
correction. Thus, the developm
techniques that minimize user i
number of algorithms have be
purpose; edge detection method
and watershed transformation [
However, the majority of t
information about the expected
some sort of user interaction
purpose of the proposed system
diagnosis process. Therefore
combining two self-organizing
hierarchical self-organizing ma
s CT Images
s
-kholy
*
and Peter Norburn*
kholy@hotmail.com
he bone structure (skull bone) is
near the sinus areas close to the
g skull bone appears bright and
pears dark and textured. Finally,
kground that leads to gaps in the
In the image domain, these
four challenging areas for
figure 1): (i) gaps in the cortical
nus areas); (ii) weak or diffused
partial volume effect in CT
orresponding to the spongy bone
and tissue-like intensities and
ons which tend to be diffused.
edges in CT image.
edical experts to segment these
manual seeding, e.g., using the
lar tools, this process is labor
e required for accurate manual
ment and use of segmentation
interaction is highly desirable. A
een invoked and tested for this
ds [2,3], active contours [4,5,6],
7,8].
the above methods require prior
features of the image and hence
n is required. This defies the
m which aims at automating the
, a new improved technique
g maps (SOM), also known as
ap (HSOM), has been developed
978-1-4244-6493-7/10/$26.00 ©2010 IEEE 290