Biomedical Engineering Reference
In-Depth Information
Finally, the contours of those regions are used to create the potential for the de-
formable models evolution. At the end of this chain the deformable model is laid
to deform under this potential. Figure 9 shows the usual pipeline for segmentation
before using the deformable model.
In simple applications where the structures are clearly defined, the first and
second stage of the process can be omitted. However, when the problem is more
complex the information obtained from graylevel images cannot suffice to define
the regions or contours of interest. It is in these difficult tasks where the full chain
must be used. The goal of the first stage, the feature extraction process, is to
provide alternate descriptions of the image according to different criteria, e.g., it
can be as simple as the variance of the neighboring pixels or as complex as the co-
occurrence matrix measures to characterize textures. As a result of applying these
different description criteria we obtain a set of
N
features,
x
=
{
x
1
,...,x
N
}
, for
each pixel in the image.
The features define multiple views of the information contained in the image.
These features do not usually define explicitly the region or contours we are looking
for as they are. Hence, in order to exploit the new information obtained by the
feature extraction process, a machine learning procedure defining which of the
areas of interest are needed. This process aims at learning of the rules that allow
the discrimination of what we have defined to be useful. In this chapter we will
look at this classification stage as a black box in which feature vectors are provided
as an input and the box returns a value indicating the belonging to one class or
to another. In fact, this value can be just a discrete boolean value that points out
if a pixel belongs to the region of interest or a likelihood value that measures the
degree of belonging to the area of interest. This difference is important because
many deformable model algorithms are defined for clear binary contours. This
fact implies that if the classifier returns a likelihood or confidence value a further
thresholding process is needed.
At this point a binary image is obtained and the contour detection is straight-
forward. The last stage of the pipeline consists of the building of the potential
field. Most deformable model techniques are applied afterward.
As we have explained, the deformable model process is the last stage in a com-
plex chain. The goal of this process is usually to give continuity and smoothness to
the scattered contours obtained in the former stages. In this sense the result of the
deformable model can be easily predicted as it is unrelated to the rest. On the other
hand, the complexity of the feature extraction and classification process and the
parameters involved in such process is usually high. This compartmentalization
of tasks is the most commonly used procedure in segmentation tasks involving
deformable models.
However, since each stage is unrelated to the others, and in particular, the
classification and the deformable model evolution stages, the control of the whole
process is difficult to adjust and usually does not generalize well — it is not easily
transferred to other environments or can easily degrade its performance if the test
