Design and Implementation - Segmentation of Hand Bone for Bone Age Assessment - page 75

each block because the context of each block is different. This information can
be acquired through a design of a series of continual decision steps in the con-
text of hand bone. The information obtained from the analysis is used to determine
whether the size of the block is suitable and therefore should remain or it is unsuit-
able and therefore should be altered. More details on blocks' size are discussed
later.
One major drawback of adopting adaptive block (sub-image) segmentation is
the choice of the block. The effectiveness of the adaptive segmentation decreases
as block's size increases due to less influence on non-uniform illumination.
However, if the size is smaller than it should be, then the content of the block
might not contain adequate information for accurate segmentation and further-
more, it is more prone to noise influence. Therefore, choosing a suitable block's
size is the main challenge of adopting adaptive segmentation, not to mention
choosing the suitable size automatically. To the best of our knowledge, there is not
yet any analytically proven solution to tackle this problem.
We suggested that the size of adaptive division block should depend on the per-
formance of clustering algorithm used in separating the bone image into compact
bone, spongy bone, soft-tissue and background by relying on the result of cluster
properties to ensure the resultant pixel intensity clusters from unsupervised clus-
tering technique are meaningful and useful in subsequent procedures of segmenta-
tion. Meaningful clusters refer to group of resultant clusters that are able to reveal
the natural structure of the data.
In the context of hand bone image processing, the structures of the data repre-
sent the anatomical structure's region of the hand bone. This process of analyzing
the properties of the resultant cluster to determine the next step of segmentation
procedure can be perceived as an intelligence of the algorithm in assessing the cur-
rent condition and then making appropriate decision to optimize the result (ful-
filment of P8). If the resultant clusters are considered meaningless after analysis,
then this result indicates that the current size of the adaptive division block is not
suitable and in turn it indicates that there is better option in terms of the size of the
block and hence appropriate adjustment is needed. On the other hand, if the result-
ant clusters are considered meaningful after the analysis, then this result indicates
that the current size is acceptable and it is permitted to proceed to next procedure
of segmentation. We propose that a similar division process is performed in those
meaningless blocks and this process repeats until all current blocks have achieved
suitable size. Next we discuss the reason why 'suitable' is used to describe the size
rather than 'optimized'.
The reason of not adjusting the size until it optimizes the objective function of
cluster validity is that being too accurate is computationally infeasible and thus it
would affect its practical usage especially in real-time computational environment
when dealing with high definition radiographs or when it is implemented in any
usage where temporal factor is an issue; for example, the implementation in hand
bone segmentation in computer-aided skeletal age scoring system. One of contri-
butions in the proposed block division scheme is that even relatively high accu-
racy of clustering can be achieved without sacrificing the computational feasibility