Result and Discussion - Segmentation of Hand Bone for Bone Age Assessment - page 113

4.3.2 Segmentation Accuracy
It is useful to evaluate the accuracy of the proposed segmentation framework
on hand bone segmentation. The purpose is to examine, despite having less
user-specified parameters, in a newly designed framework, can the accuracy of
segmentation achieve the similar level to AAM. The accuracy of AAM though
is difficult to be measured as there are many undefined parameters in literatures
of hand bone segmentation using this method. Secondly, it is obvious from the
theoretical aspects of AAM that only if it is given sufficient number of train-
ing images that encompasses sufficient expected changes of hand bone number,
locations and sizes, all parameters are correctly tuned, hand bone anatomical
structures of all training radiographs are accurately delineated, only then the
segmentation can be almost perfect. Note that not all segmentation methods can
hit the accuracy of being 'almost perfect' even if the parameters are given the
best tunings; for example, the global thresholding method discussed in Chap. 2 ;
even the 'best' threshold is chosen, the threshold is unable to overcome the une-
ven luminance and the inherent intensity distribution of hand bone radiographs.
Therefore, it was the objective of this sub-section to justify the accuracy of the
segmentation framework, to examine its 'best' accuracy, despite having less user-
specified parameters.
There are multitude segmentation techniques and frameworks exist in litera-
ture [ 3 , 7 , 8 ]. However, difficulties arose if the performance of those developed
segmentation methods in literature are to be measured. It is claimed that the
unsupervised segmentation evaluation has an edge over other type of segmenta-
tion evaluation such as unsupervised segmentation evaluation and human visual
inspection segmentation evaluation. The key advantage is its self-tuning potential
since no reference image is needed as in supervised evaluation. Besides unsuper-
vised evaluation, there is another common group of methods in segmentation eval-
uation which can be categorized as supervised evaluation. Supervised evaluation is
a comparison made against a predetermined manually segmented reference image.
Similar to human inspection method, supervised evaluation involves human assis-
tance in the early stage such as segmenting a reference image, which could be
even more tedious than human visual inspection alone, especially in the context of
evaluating hand bone segmentation; this is due to the inherent countless variations
in number and size of ossification sites along the hand bone development.
Attributable to the abovementioned reasoning, unsupervised methods have been
adopted in this topic as performance evaluation to assess some of the result such
as in evaluating the abilities of histogram equalization, firstly, in solving the prob-
lems of over-enhancement, detail and brightness preservation as well as the con-
trast enhancement; secondly, in evaluating the ability of anisotropic diffusion in
solving the problem of high diversity in pixel intensity in hand bone radiograph.
However, unsupervised evaluation is not adopted in evaluating the segmentation
result in hand bone segmentation. Unsupervised evaluation of segmentation in lit-
eratures, though have an edge in term of time taken and objectivity over human