Result and Discussion - Segmentation of Hand Bone for Bone Age Assessment - page 111

problem is that N SP always requires a substantial numbers so that to sufficiently
describe the shape of the hand bone, for example, in [ 4 ], the author utilized 64
shape points to delineate the hand bone shape. On the contrary, in the fuzzy quad-
ruple division scheme, or any fuzzy inference system, the number of membership
functions is averagely around 3-6 in order to sufficiently describe the expert vague
knowledge. This explanation goes similarly to the training examples; in [ 4 ], the
author utilized 1,559 samples to train the model so that the model could capture
the changes of shapes existed in the samples. However, in the proposed segmenta-
tion framework, no training sample is required. The large numbers of shape points
and training samples magnify the constants in the Eq. ( 4.5 ) as shown in Eq. ( 4.6 ).
In contrast, it is compared to Eq. ( 4.7 ), which is the numerical representation of
Eq. ( 4.3 ):

(4.6)

N TNP = ( 1559 ∗ 4 ) ( 64 ) + 10 = 3,99,112

(4.7)

N TNP = 4 + 3 [ ( 2 ) + ( 2 ) + ( 3 ) + ( 3 ) ] + 8 = 42

The above illustration explains the reason for the large number of difference
between the N TNP of AAM and the proposed segmentation framework. From the
comparisons and arguments above, the impact of the user-specified parameters for
both AAM and the proposed segmentation framework in hand bone segmentation
in following paragraphs is interpreted.
The proposed segmentation framework requires relatively much less numbers
of user-specified parameters if it is compared to AAM framework. The above
fact can be viewed from a few perspectives. First of all, from the perspective of
automaticity, large number of user-specified parameters simply means that it
needs human decisions and labors to perform the task; the higher the number of
user-specified parameters indicates the higher the difficulty in transforming those
parameters becoming automated. From the perspective of the time taken in seg-
mentation task, the high number of user-specified parameters indicates high time
consumption as each parameters needs to be adjusted. From the perspective of
adaptability, higher number of user-specified parameters can indicate higher prac-
tical flexibility and adjustment because the operator possesses powerful cognitive
ability relative to computational algorithm. From the perspective of noise resist-
ance, higher number of user-specified parameters in the context of AAM in hand
bone segmentation can be considered as being inherently more robust because the
trained model has been placed very close to the expected location of each struc-
ture in hand bone and thus the trained model is less susceptible to the noises in
soft-tissue region and background (the proposed segmentation framework has
incorporated the anisotropic diffusion to reduce the influent of noises found bone
structure, soft-tissue region and background).
The types of the parameters of both frameworks are different. Thus, the natures
of those parameters are different as well. The user-specified parameters in proposed
segmentation framework can be easily replaced by simple analysis or assumptions
such as in the relative importance of the MBOBHE, the relative importance of the
three properties of histogram equalization can be assumed as being equal. Even if