Result and Discussion - Segmentation of Hand Bone for Bone Age Assessment - page 102

(2) The second experiment was set up to justify the positive effects of adopting
anisotropic diffusion in the segmentation framework compared to conven-
tional filtering methods in tackling the problem of inherent high variations
within the bone structures.
(3) The third experiments consisted of analytical evaluation and empirical evalu-
ation was set up to justify the performance of the proposed segmentation
framework.
4.2 Anisotropic Diffusion in the Proposed Segmentation
Framework
In previous chapter, the details of anisotropic diffusion and its role in the segmen-
tation framework have been explained. In this chapter, the effect of anisotropic
diffusion on hand bone radiographs is justified. As shown in the Fig. 4.1 , the bone
area had been smoothed to become homogenous area; the black holes and dots in
had been filled by similar pixel intensity with the surrounding bone. Despite this
filtering process, the edges of hand structure are preserved and can be clearly seen.
Figure 4.2 illustrated the diffusion effects of various filtering methods:
Gaussian filter, average filter, wiener filter [ 1 ] and Symmetric Nearest Neighbor
(SNN) filter [ 2 ] and anisotropic diffusion. From this figure, it is observable that the
effect of anisotropic diffusion was more favorable in comparison to other diffusion
methods as the edges had been preserved and even enhanced while the degree of
heterogeneity within the hand bones has been reduced by eliminating the random
noise. Unlike anisotropic diffusion, the other filters were not adaptive to the infor-
mation contained in the moving kernel. In other words, the filters imposed identi-
cal degree of diffusion and identical direction of diffusion to every single pixel
without considering the suitability of information resided in the kernel. Therefore,
as expected, the edges are blurred.
To quantify the changes in homogeneity, an experiment is conducted to evalu-
ate the homogeneity of 12 radiographs from each age group after being diffused
by anisotropic diffusion. Table 4.1 showed the experimental result of expected
image homogeneity before and after the anisotropic diffusion processing. From the
table, it is observable that the homogeneities after the diffusion are consistently
larger the image homogeneities before anisotropic diffusion over all age groups.
The increase of homogeneities is important to subsequent steps in the proposed
framework's sub-routine. The homogeneities blend together the noises and smooth
the uneven texture of bones so that the exhibited histograms of bone regions are
more prone to having separated 'hills' that are inherently easier to be distinguished
by relatively simple clustering or thresholding techniques.
As a whole, anisotropic diffusion has an edge over the traditional scale-space
filtering methods in terms of the relatively low complexity in computation. This advan-
tage capable to extend its applicability of anisotropic diffusion for general purposes;
anisotropic diffusion takes structures outlines into consideration in filtering; therefore,