Design and Implementation - Segmentation of Hand Bone for Bone Age Assessment - page 74

Eq. ( 3.52 ) to form the final segmented image. The maximum number of block
depends on the quadruple division algorithm in Sect. 3.4.2 .

max i

(3.52)

Final Segmented Image = B ( i , k )

i = 1

Next we have to optimize the result by instilling some rules and intelligence for
the machine incorporating with homogeneity to adaptively opt for the suitable size
of block in every region using the proposed automated block division scheme as
explained in following sections.

3.4.2 Automated Block Division Scheme in Adaptive
Segmentation

In designing the adaptive clustering algorithm, two decisions are crucial: Firstly,
the size of the block and secondly the robustness of the designed clustering algo-
rithm under different circumstances. The common desired properties of choosing
the blocks' size and designing clustering algorithm are automated (fulfilment of
P6) and adaptive (fulfilment of P10). Being automated implies that there is no
involvement of user-specified parameters or any manual manipulation to avoid the
drawback of subjectivity and inconsistency of making decision by users. Being
adaptive indicates that each decision is made according to its given context on the
ground that the quantities and characteristic involved are dynamic and should not
be treated similarly using single preset parameter value so that it is adaptive to dif-
ferent environment or context.
The conventional unsupervised clustering fulfils the requirement of being auto-
mated, but not the second requirement which is robust under different conditions.
The inconsistent robustness of the conventional clustering algorithm is attributable
to various factors: the number and 'location' of initial centroids, relatively unequal
proportional size, densities of the natural structure of data, the shape of the natural
structure of data, the disturbance of outliers.
In the context of hand bone x-ray radiograph, the main problem is the unpre-
dictable proportional number of pixels in each of the underlying natural parti-
tions of background, soft-tissue region, cancellous bone and cortical bone while
performing the adaptive segmentation. For example, the number of pixel belong
to cortical bone region could not be predicted in some block where most of the
region in the corresponding block are background and soft-tissue region. In this
kind of block, the conventional unsupervised clustering might fail to optimize the
objective function and lead to the undesired visual effect after partitioning the pix-
els inside the block into four clusters. Therefore, there is a need to analyze the
proportional number of pixels in each of the four partitions before processing
them using unsupervised clustering. In other words, extra information about the
block is required to have optimized performance in the unsupervised clustering of