Literature Reviews - Segmentation of Hand Bone for Bone Age Assessment - page 20

bone radiograph has more intense illumination relative to upper region of the
radiograph. The Global thresholding neglects this uneven illumination and this
affects the segmentation result. The effects can be perceived obviously espe-
cially via Fig. 2.2 a-c.

Another critical problem of single Global thresholding is the choice of the
threshold value to obtain favorable segmentation result [ 38 ]. In fact, even the
'best' threshold value is selected, the resultant segmented image in the context of
hand bone radiograph and in other medical image processing remain inferior. This
fact is inevitable due to the nature of Global thresholding and the nature of hand
bone segmentation: only one threshold is selected for thresholding. One improve-
ment for this limitation is by adopting multiple Global thresholding [ 39 ]. Multi-
level thresholding classifies the image into multiple classes (>2) [ 40 ]. The multiple
thresholding can be represented as follows:

g 1 IF f ( x , y ) > T 1
g 2 IF T 1 < f ( x , y ) ≤ T 2
. .
g n − 1 IF T n − 2 ≤ f ( x , y )< T n − 1
g n

(2.2)

f ( x , y ) =

IF f ( x , y ) ≥ T n − 1

where g i denotes group of pixel that represents an object or a background. T i
denotes the threshold value. The f ( x , y ) denotes the image pixel intensity in 2D
gray-scale image in coordination ( x , y ) .
Multiple thresholding might solve the problem arises from the assumption that
the input image is of bi-modal type but solve not the problem arises from assump-
tion that the input image is of even illumination. In next sub-section, we would
review and examine the local thresholding that is claimed to be more effective in
tackling the problem of uneven illumination [ 41 ].

2.3.2 Local Thresholding

Local thresholding is segmentation using different thresholds in different sub-
images of input image [ 42 ]. The input image is firstly divided into a number of
sub-images, and then in each sub-image, suitable threshold is chosen to per-
form the segmentation, this process repeats until all sub-images undergo the
Thresholding segmentation. Adopting different threshold in different region of the
input image is proven to be more effective than Global thresholding in that it is
easier to obtain well-separated bi-modal or multiple-modal distributions in the sub-
images and hence it improves the segmentation result [ 43 ]. In addition, sub-images
are more likely to have uniform illumination. This implies that local thresholding
could resolve the problem that arises from the non-uniform illumination [ 44 ].