Literature Reviews - Segmentation of Hand Bone for Bone Age Assessment - page 15

2.2 Existing Hand Bone Segmentation
All the computer-aided BAA system require a segmentation stage with the purpose
of eliminating the background, noise, soft-tissue region that contains no pertinence of
information in the deduction of the skeletal maturity measurement [ 1 - 6 ]. Instead, this
undesired information will affect the subsequent stages of computer-aided skeletal age
scoring system and in turn deteriorate the result accuracy. However, most of the con-
ventional methods used in this pre-processing stage are either not practical in terms
of resources consumption in most of the context or are not yet developed into fully
automated framework. Besides, most of the researchers perform the segmentation after
obtaining the region of interest (ROI) to reduce the difficulty of segmentation [ 7 , 8 ]. In
fact, the segmentation accuracy and ROI searching ability can be enhanced by employ-
ing the algorithm after the hand bone partition from the soft-tissue region. Therefore,
As the main pre-processing stages of the computer-aided system, the output accuracy
and practicality of segmentation are critical because the eventual outcome of the com-
puter-aided skeletal age scoring system depends on this partition procedure [ 9 ].
Substantial studies have been conducted to solve the hand skeletal bone parti-
tion problem to exclude it from soft-tissue region and background. Most of the
studies involve the employment of threshold that is considered impractical in the
hand bone segmentation as the soft-tissue region possesses pixel that have similar
intensity to pixels found in spongy bone [ 1 , 3 , 10 , 11 ]. Moreover, majority of the
studies, implements the active contour model after obtaining the region-of-interest
(ROI), [ 12 ]. This leads to drawbacks such as the contour sensitive to intensity gra-
dient, depends heavily on initiate position and exhibits inability in extending into
concavity [ 13 - 15 ]. Besides, some studies have adopted the statistical analysis to
acquire the membership of each pixel i.e., to determine the labeling of the pixels
either to the bone or the soft-tissue region [ 16 , 17 ]. Also, some researchers incor-
porate various segmentation techniques into the hand skeletal bone segmentation
[ 5 , 18 ]. The summary of the studies are described in the next few paragraphs.
As early attempt, Michael and Nelson [ 19 ] proposed a computer-aided diagno-
sis (CAD) system for BAA that includes pre-processing, segmentation and meas-
urement. The image pre-processing is done via the histogram equalization and then
followed by binary conversion of the radiograph and implementing the threshold-
ing using intensity of pixels to eliminate the background by the model parameters.
By using the model parameter, the main drawback is the overlapping problem of
pixel intensity in bone and background. Moreover, high sensitivity to illumination
uniformity and the presence of soft-tissue region locating around the hand bone
have further deteriorated the result. Manos et al. [ 20 , 21 ] proposed a framework for
the automatic hand-wrist segmentation; they have implemented a region growing
and region merging technique after performing the edge detection during the pre-
processing. Amidst this technique, thresholds involve to determine the efficiency of
the edge and growing and merging algorithms. Furthermore, region growing result
depends heavily on the performance of edge detection. Lastly, the region merging
depends on grey level similarity size and connectivity which bear a risk of combin-
ing the epiphysis sites that are situated around the metaphysis.