Literature Reviews - Segmentation of Hand Bone for Bone Age Assessment - page 32

The essential concept of region-based segmentation has been reviewed. The pur-
pose is to identify coherent regions defined by pixel similarities. The main challenge
of this type of segmentation is often related to the pixel similarities: what are the
features that should be adopted as similarities measurement and how the thresholds
of chosen features should be set in defining the similarity. The selection of features
is difficult as they depend on application. For example, if the targeted object is not
a connected object, pixel intensity is not suitable as pixel similarities measurement.
The setting of threshold is another tricky challenge as it manipulates the trade-offs
in terms of flexibility. For example, if the threshold is set too low, the inferior effect
of over-segmentation occurs because pixels easily surpass the threshold leading to
larger coherent regions than the actual objects; if the threshold is set too high, oth-
erwise occurs. Region based segmentation unable to segment objects that contain
multiple disconnected regions and therefore, in the context of hand bone segmen-
tation, applying only region-based segmentation is inappropriate as children hand
bone of different ages for BAA involve different number of bones regions.
2.6 Hybrid-Based
As previously discussed, the Thresholding segmentation, edge-based segmentation
and region-based segmentation have their own merits and each in their own way
has made an important contribution. Therefore, it is assumed that if all these attrib-
utes together are combined, a better segmentation result can be obtained. In the
following, watershed algorithm based on embodying the basic concepts of previ-
ous discussed thresholding, region growing and edge detection, is discussed. The
main concept is to interpret the image topographically using 3D visualization on
the image which refers to the two spatial coordinates and pixel intensity.
2.6.1 Watershed Segmentation
Watershed Segmentation is a hybrid-based segmentation that bases on math-
ematical morphology. An input image in watershed Segmentation is represented
topographically as spatially distributed terrain altitudes in accordance to the
numerical value of pixel intensity. Such image representation is useful to visual-
ize the notions of watershed as light area, catchment basins and minima as dark
areas. The idea of using watershed transformation is first introduced by Digabel
and Lantuéjou [ 82 ] using binary image. It was then extended to grayscale images
based on immersion analogy by Beucher and Lantuejoul [ 83 ]. Further study about
the practical and algorithmical aspects of watersheds when it combines morpho-
logical tools are performed by Vincent and Soille [ 84 ] to serve as the basis of
modern watershed Segmentation algorithm. The watershed Segmentation in fol-
lowing paragraphs are described briefly [ 33 ].