Literature Reviews - Segmentation of Hand Bone for Bone Age Assessment - page 40

in Eq. ( 2.13 ). This equation represents the operations of scaling the point ( x , y ) by
s , rotating the point by ( x , y ) and translating point ( x , y ) by

. The flowchart of

x t , y t

the process is shown in Fig. 2.9 .

s COS (θ) − s SIN (θ)
s SIN (θ) s COS (θ)

x
y

x t
y t

x
y

= + (2.13)

T

(b) Profile Model

The profile model defines the desired image information around the landmarks
region. This information is utilized to model the expected image characteristic
around each landmark; the expected characteristic is modeled through various
training examples during the training phase; the most commonly adopted charac-
teristic is the edge strength of the targeted image object's boundary; the landmark
examines the edge strength along normal to the object boundary edge and moves
toward the strongest object boundary edge. Generally, during the matching phase,
the landmark adjusts its current position and makes displacement around its sur-
rounding region to search for the optimized position accordingly to the profile
model (these adjustments are constrained by shape model which will be discussed
later). After every landmark moves to its temporary position accordingly, the over-
all shape formed by connecting these landmarks is known as the suggested shape.
This process of searching and adjusting to form temporary shape iterates until con-
vergence to obtain the final shape.

(c) Shape Model

Shape model, as mentioned above, limits the adjustments of landmarks. The
purpose is to prevent deviations on the landmarks adjustments which would result
in forming an allowable overall shape. With the model constraints, the final shape
remains resemble to training shapes. This idea of global shape control is the main
distinction between active contour model and active shape model. This shape
model, similarly to profile model, is obtained across training shapes. The detail of
the calculations and computational steps can be found in Cootes et al. [ 97 ].
In short, the final shape is determined by both the profile model and shape
model: the profile model inspects landmarks surrounding and informs the landmarks
about the movement suppose to make in order to fulfil as close as possible to the
predefined profile; the shape model inspects the shape as a whole and informs land-
mark about the permissible movement while trying to fulfil the predefined profile.
ASM has been applied by Thodberg and Rosholm [ 30 ] to address the prob-
lem of hand bone segmentation. Extensive training that have to be done to com-
plete the model in order to imitate the recognition understanding of human beings
in segmenting the hand bone. Note that the initiation of set points placement to
mark the spatial position of hand bone shape demands expert to be the opera-
tors. Both requirements of training set and human expert are the main weakness
of this model in addressing the problem. It is tedious, subjective and time con-
suming to delineate the shape from a large training set, not to mention the critical
issue of the availability of these resources. Therefore, an alternative segmentation