Biomedical Engineering Reference
In-Depth Information
shape is chosen; how the texture is represented; how the texture is modeled. It
is crucial to improve this discriminative ability to perform segmentation tasks
effectively.
3. Inconsistent robustness under different circumstances: the performance of the
system is influenced by different conditions such as the existence of pose varia-
tions, uneven illumination, and absence of features, low resolution and the pres-
ence of noises.
AAM is a very useful model as it can capture the mode of variations of deform-
able objects given a set of training examples. The mode of variations includes
shape and texture as a whole. Besides, it can perform the projection of object onto
low dimensional subspace to reduce redundancy and capture main component of
variations. Thus, it has been implemented in a lot of applications especially medi-
cal image segmentation. Nonetheless, it has limitation in efficiency, discriminative
ability, and robustness in different condition. In the problem of hand bone seg-
mentation, the same group of researchers that adopted ASM has extended their
works by applying AAM [ 29 , 116 ]. The weaknesses discussed in applying ASM
remains because the technical differences between AAM and ASM enhance only
the robustness in terms of prior knowledge and the information around the object
that have been incorporated into the model, not the practicality in terms of avail-
ability of training set and expert operators.
The deformable models, ASM and AAM, undoubtedly are powerful segmenta-
tion methods. However, they are not without weaknesses and are not best method
for automated hand segmentation. The reasons are summarized as follow:
1. The landmarks placement has to be manually annotated by users. Incorrect
landmarks placements lead to unreliable capture of shape variability.
2. The number of landmarks has to be specified by user manually. Insufficient
landmarks lead to failure in obtaining the shape of the targeted structures;
excessive landmarks lead to computational inefficiency.
3. The training phase requires a lot of training examples in database which is not
necessarily available in many applications. Insufficient training examples lead
to failure in generalizing the mean structure's shape.
4. The nature of hand bone development of children: different number and size of
bones in different range of age complicated the implementation ASM and AAM
especially in establishing the general form of mean shape.
5. The alignment phase is uncertain in terms of its numerical stability: the con-
vergence of the mean model in the iterative method has not been devised math-
ematically and prone to errors.
6. The choice in retaining the number of eigenvectors in principal component
analysis has to be determined correctly by user. Incorrect decision leads to fail-
ure in capturing the representative points of the shape; consequently, inaccurate
model is constructed and leads to undesired segmentation result.
7. Variations in hand structural positions are often largely deviated and this
devotes to non-linear parameter relations that invariably impede the accurate
segmentation as a whole.
Search WWH ::




Custom Search