Result and Discussion - Segmentation of Hand Bone for Bone Age Assessment - page 114

visual system, they lack of background knowledge as in human and they are too
general to be applied in such a specific system such as computer-aided skeletal age
scoring system which requires evaluator to distinguish which features are pertinent
in segmented image that determines an accurate result of bone age assessment.
A so-called 'good' segmentation evaluation is a relative quality. Whether it
is favorable should depend on the application. As each application has different
focus and hence different criteria should be categorized as a favorable resultant
segmentation. In the context of bone age assessment, the focus is on the features
which will affect the score in bone maturity such as the thickened white line along
the border, continuous border of epiphyseal center, the diameter ratio of the epi-
physeal centre to the metaphysis, the dark line of cartilage, distinct thickened
proximal border and etc. However, to the best of our knowledge, the existing
unsupervised segmentation evaluation methods are not tailored for hand bone seg-
mentation in the application of computerized TW3 bone age assessment. Using a
general segmentation evaluation is inappropriate and inaccurate to reveal the real
performance of each segmentation method.
Due to the above reasons, a specially tailored human visual based evalua-
tion method for hand bone segmentation in the context of bone age assessment
is designed and used to evaluate the segmentation result. This type of supervised
evaluation can be categorized as subjective evaluation as well, which is claimed
to be the most common evaluation methods. The major disadvantage of this sub-
jective evaluation is that it is very subjective, as its name suggests. However, if
the reference object such as hand bone is manually segmented with high accuracy
and the relative contribution of each feature are set correctly, and then the evalua-
tion can be regarded as considerably objective. Besides, it is very time-consuming
if the experiments involve a large number of data. However, in the case of this
topic, such evaluation is the best evaluation compared to other existing supervised
and unsupervised evaluation method. Therefore, to maximize its accuracy of eval-
uation and minimize the bias caused by subjective evaluation, a large set of test
images from the database are adopted, and four evaluators who have been inspect-
ing more than 1,500 hand bone radiographs for manual TW3 bone age assessment
are given task to evaluate the result. Large set of test images and multiple eval-
uators are effective in reducing the undesired bias such as subjectivity. Another
important factor is a set of well-designed, clear and application-related guidelines
in the evaluation given to the evaluators.
A set of good guidelines should be related to the application or the purpose
of the segmentation. In the context of bone age assessment, the key point is that
if less concern is on evaluating how accurate the segmented image resemblance
the desired segmented image for the bone areas, but how much pertinent informa-
tion has been retained in the segmented image. An example is shown in Fig. 4.5
to illustrate this concept. Both Fig. 4.5 a, b have same amount of pertinent infor-
mation as the regions-of-interest (regions within red rectangular) are not affected
by the artifacts, or in other words, the artifacts (region within blue rectangular)
of Fig. 4.5 b occur in non-region-of-interest. In terms of information preservation,
both images are identical, in the context of TW3 bone age assessment.