Literature Reviews - Segmentation of Hand Bone for Bone Age Assessment - page 38

purpose of this process is to build a typical targeted shape by analyzing various
shapes variations over the training images. The process begins with placing land-
marks to define the shapes, followed by aligning the shapes to standardized form,
and finally, build a statistical model based on the parameter.
'Landmarks' are points that best describe the shape of the object. The process
of placing the landmark is a labeling procedure. 'Landmarks' placement can be
perceived as feature extraction. These points serve a purpose: as input to the sub-
sequent algorithm to represent the shape of the object. The selections of locations
for these points are crucial. The common guidelines while placing the points or
'landmarks' are as follow:

1. Place the points on the objects boundaries.
2. Place the points evenly along the boundaries.
3. Place the points on the same location in every training image.
4. Place the points on corners or high curvature edges of objects boundaries.
5. Place the points on 'T' junctions between boundaries.
6. Place intermediate points between main landmarks.

After placing the 'landmarks', all the 'landmarks' coordinates are collected and
vectorized as an n-point column matrix, X, as presented in equation

(2.12)

X = ( x 1 , x 2 , ... , x n − 1 , x n , y 1 , y 2 , ... , y n − 1 , y n ) T

The users have to specify the number of 'landmarks' and the location of 'land-
marks'. These choices are of essence; incorrect decisions are likely lead to disas-
trous result in latter stages.
The main stages of active shape model are explained briefly as follow:

(a) Model Landmarks Placement

After defining the 'landmarks', they have to be further processed. The process-
ing of these points is known as alignment, a filtering procedure of translational,
rotational and scale effect of the shape. This process aligns the landmarks/shape
by using training images in order to cope with the non-standard shapes (differ-
ent orientation, size and position) in training images. The flowchart summarizing
alignment procedure is shown in Fig. 2.9 . The purpose of aligning the shapes in
training images is to assure that all the shapes are invariant in terms of orientation,
size and position so that the true shape representation is obtained. These invariants
are vital to enable comparison among different normalized shapes. In other words,
the shapes in test images require transformation before the training process begins:
the size invariant is achieved by scaling operation whereas the operations of rota-
tion and translation deal with orientation invariant and position invariant respec-
tively. This shape invariant is to ensure that all shapes in training images possess
approximately same position and same size in x-y plane, as well as to ensure that
all shapes are different by minimum distance; only then the shapes in test images
are ready to be learnt by the subsequent algorithms.
The Scaling, translation and rotation transformation are known as similarity
transformation. The similarity transformation can be expressed mathematically as