Literature Reviews - Segmentation of Hand Bone for Bone Age Assessment - page 41

framework has to be established when the resources are limited and this motivates
the research of this topic.
2.7.3 Active Appearance Model (AAM)
AAM is a statistical model of shape and grey-level appearance of the targeted
object proposed by Edwards et al. [ 108 ]. The final aim is to generalize the model
to all valid example [ 109 ]. The relationship between the model parameter dis-
placements and the errors between training example and a model instance is
learned during the training phase [ 110 ]. By computing the errors of fitting and
using the previously obtained parameters, the current parameters with the intent of
improving the current fitting can be updated.
AAM and the closely related to the concepts found in the methods of Active
shape model. The AAMs is most frequently being adopted in the application related
to face modeling. Besides face modeling [ 111 ], it has been implemented in other
applications as well such as in medical image processing [ 112 - 114 ]. The typical
first step of AAM is to fit the AAM to an input image using model parameters that
maximize the matching criteria between the model instance and the input image.
The model parameters are then passed to a classifier to yield t classification tasks.
Fitting the AAM to an input image involves solving a non-linear optimiza-
tion problem. The conventional method of solving the problem is by updating the
parameters iteratively. This update has to be incremental additive and the param-
eters refer to shape and appearance coefficients. The input image can be warped
onto the model coordinate frame by using the current shape parameters estima-
tions. The error between the model instance and the fitting of AAM onto the image
can be computed. This error then acts as feedback in next iteration that would
affect the updates of the parameters. The constant coefficients in this linear rela-
tionship between the updates and errors can then be found either by linear regres-
sion or by other numerical methods.
Although the AAM appears to be the useful model-based system in medical
image segmentation, it has constraints that impede its performance in practical
application [ 115 ].
1. Low efficiency in real-time systems: current algorithm of AAM consumes a lot of
time and space computational costs. Thus, it is of prime importance to minimize
the complexities in time and space needed to perform the algorithm in order to
realize it in real-time system. The efficiency is mainly affected by the following
factors: manual landmarks placement, complex texture representation in high res-
olution medical image, iterative procedure in solving the optimization problem.
2. Low discriminative ability for recognition and segmentation systems: only
a group of object is being modeled and thus it is considered as a generative
model which possesses no ability to classify different objects. This ability
depends on the accuracy of model fitting which are affected by how the prior