Biomedical Engineering Reference
In-Depth Information
Figure 4.
Filter pipeline for STOP and GO active models.
data change a little. Therefore, it looks reasonable that the mixture of classification
and evolution could help to improve the control of the processes involved.
In this trend, the most well-known techniques are the region based scheme.
However, the main drawback of these techniques is that, although it includes
the classification task in its evolutive model, the resulting segmentation does not
differ from a batch approximation. Note that the boundaries used in the region
competition are those defined by the Bayessian rule:
P
(x
|
c
=
A
)
P
(
c
=
A
)=
P
(
x
c
=
B
)
P
(
c
=
B
).
From the point of view of the segmentation pipeline, what we propose is
a different embedding of the classification task in the evolutive model. In our
approach, both tasks are performed at the same time. This means that regions are
not
a priori
labeled. On the other hand, we just need a confidence rate of the region
belonging to the area of interest. In this sense we observe three main scenarios
in which this confidence rate can be obtained: a filter approach, a likelihood map
approach and a classification confidence rate approach. The following sections
are devoted to the description of these scenarios and possible tools for enhancing
the responses if they are not well defined.
|
4.1. Filter Potentials
Filters are well known tools for enhancing or extracting information from
images. Usually, the result of the filtering process of an image shows high values
in the areas where the filter has been designed to work, i.e., if we are using a contour
detector filter, the values of the filtering process will be higher where the contours
are more likely to be found. However, not all filter designs are meant to work in this
way, i.e., isotropic filtering just modifies the gray values according to a diffusion
criterion. In this section, when we talk about filter potential we just consider those
filters in which the premise of having high values at the structures of interest holds.
These filters are the ones usually used in feature extraction processes.
Figure 4 shows the pipeline before the deformable model. As one can observe,
the response of the filter is used directly as an input to the deformable model. The
only minor question is that the filter response has to be normalized between 0
and 1.
