Literature Reviews - Segmentation of Hand Bone for Bone Age Assessment - page 31

where n denotes total number of regions, R if . Where N denotes the immediate

neighboring pixels of each element. This indicates that a pixel can only be considered
as the neighboring pixel of a region if it fulfils the requirement that its neighboring
pixel is overlapping with the region and at the same time it is not yet being allocated.
The expansion depends on a defined similarity measure S ( I , J ) where I and J
are adjacent pixel. This similarity measure is designed such that the score is higher
if both I and J have common desired image characteristics and thus they should be
grouped together; this process iterates until convergence. For instance, if the com-
mon characteristic is intensity value, the membership determination in 2D case
should be set as follows where T denotes the threshold:

( x , y )

M ( x , y ) ∈ R if if S ( M ( x , y )) ≤ T

(2.8)

In comparison with deformable model based segmentation, region based seg-
mentation is considered relatively fast in terms of computational speed and
resources. Besides, it is certain that segmentation output is a coherent region with
connected edges. Simplicity in terms of concept and procedures is an advantage of
region growing for immediate implementation [ 78 ].
Region based segmentation is insensitive to image semantics; it does not rec-
ognize object but only predefined membership function [ 79 ]. Besides, the design
of the region membership is as difficult as setting a threshold value; region based
segmentation is unable to separate multiple disconnected objects simultaneously.
The assumption that the region within a group of object is homogenous has low
practical value in hand bone segmentation due to the fact that the bone is formed
by cancellous bone and cortical bone that has high variations on texture and inten-
sity range. Besides, in the presence of noise or any unexpected variations, region
growing leads to holes or extra-segmented region in the resultant segmented
region and thus has low accuracy in certain condition [ 80 ]. The number and the
location of seeds and membership function in seeded region growing, as well as
the merging criteria in split-merge region growing, which will be discussed later,
depends on human decisions which is subjective and laborious.

2.5.2 Region Splitting and Merging

Another famous region growing methods is the split and merge algorithm; split
and merge is an algorithm splitting the image successively until a specified num-
ber of regions remain [ 81 ]. To perform the split and merge region growing algo-
rithm, firstly, the entire image is considered within one region. Then the splitting
process begins in the region in accordance to the homogeneity criterion; if the
criterion is met, then it splits [ 33 ]. This splitting process repeats until all regions
are homogenous. After the splitting process, the merging process begins. Initially,
comparison among neighborhood regions is performed. Then, the region merges
to each other according to some criterion such as the pixels' intensity value where
regions that are less than the standard deviation are considered homogenous.