Literature Reviews - Segmentation of Hand Bone for Bone Age Assessment - page 33

Three group of point need to be defined to understand the algorithm. First
group of points refers to points that form the regional minimum. The second group
of points, namely the 'catchment basin' refers to points that if a drop of water fall
on them, the drop of water will certainly fall to a single minimum. The third group
of points, namely the 'watershed lines', refers to points that if a drop of water fall
on them, the drop of water will fall with equal likelihood to more than a single
minimum. Usually the third group of points forms lines on the topographic surface
as ridges to delineate the catchment basin [ 85 ].
The segmentation task can be accomplished if the watershed lines can be
found. The idea of searching the watershed line is to imagine, given an image,
which can be visualized as a topologic surface, and water is added to flood the
topologic surface at uniform rate. The water level keeps rising until the water is
about to fully cover the catchment basin, then imagines a dam is built to prevent
the water from merging the catchment basin. This process repeats until only all the
top of the dams of all the catchment basin remain above the water level. These top
of dams form boundaries which are watershed lines. These lines are the desired
result of watershed Segmentation algorithm.
The classic watershed Segmentation encounters problem of over-segmentation
stems from spurious minima. Therefore, various improvement strategies have
been published in terms of the algorithm efficiency [ 86 ], incorporation of other
techniques such as multi-resolution [ 87 , 88 ], wavelet analysis [ 89 ] or both [ 90 ],
utilization of prior information such as shape and appearance [ 91 ], combination
with artificial neural network [ 92 ]. Despite various improvements, the water-
shed remains corresponding to image gradients that are always affected by noises
within the image that lead to spurious edges. This in turn leads to over-segmenta-
tion. Besides the watershed algorithm require threshold setting to determine the
object gradients. This thresholding is virtually of little avail in hand segmenta-
tion due to the uneven illumination and various gradient definitions for anatomi-
cal structure edges. In the context of automated hand bone segmentation, Fig.
2.7 demonstrates that the watershed algorithm is not robust enough to solve the
problem of noises, undistinguishable anatomical structures and uneven illumina-
tion. The reasons stem from the inherent property of watershed Segmentation as
it requires thresholding in converting the input image into binary image or gradi-
ent image as pre-processing [ 93 , 94 ]. Most importantly, the hand bone for bone
age assessment encompassing different age group consists of multiple numbers of
bones and this objects inherent property simply fails any type of watershed algo-
rithms. The result in Fig. 2.7 illustrates that watershed algorithm is not efficient in
segmenting the hand bone.
Watershed Segmentation has been reviewed. This technique resolves some
inferior effect of conventional edge-based segmentation: dependency on gradi-
ent, resultant thick lines, detected edges pixels are far apart. Unfortunately, this
method is too sensitive towards noise and irrelevant changes in pixel intensity
leading to over-segmentation. However, this drawback can be greatly improved
by using markers found by features associated with background or object in
watershed-processed image. Besides, this algorithm usually processes edge