Biomedical Engineering Reference
In-Depth Information
However, these approaches require a previous classification or contain implicit
classification schemes. In this sense, little control over the false positive and false
negative regions is obtained.
Considering the general problem of region-based segmentation, we proposed
in [8] a new geodesic snake formulation where the terms ruling convergence and
regularity are decoupled. As a result, the curvature term is restricted to the shape
regularity at the last stages of the snake deformation. Furthermore, any vector
field properly defining the path-to-target contours or regions of interest is suitable
to guide the model. In our formulation, we also split the two main motion terms
guiding the deformable model to its goal. On one hand, we have the external
attractor vector field (GO term), which guides any external curve to the regions
of interest. And, on the other hand, a repulsive vector field (STOP term) is used
to cancel the forward motion of the GO term at the borders of interest. Due
to these two terms, we call this technique the
STOP and GO
active models. In
order to use the decoupling strategy, a characteristic function is needed. The
characteristic function is defined with the value 1 where the regions of interest
are located and 0 in the remainder. To sum up, a mask defining the object of
interest would be the ideal tool to bound the scope of the curvature term and
perform any decoupling. However, in real applications we do not have this mask.
To address this issue, a substitute estimation must be provided. This point is
one of the key issues of this formulation. In this chapter we explain a powerful
technique that allows any approximation to a mask to be used. In this sense, this
technique increases the number of possible definitions of the regions of interest.
Therefore, not only the classical region-based definitions or contour images are
suitable for deformation of the snakes, but any map ranging from filter responses
to likelihood or confidence rate maps can be used. This chapter explains in detail
how these alternative deformation spaces are designed and embedded into the
deformable model equation. We show results applied to two different medical
images: intravascular ultrasound images (IVUS) and intestinal capsule endoscopy.
Intravascular ultrasound is an image modality based on the ultrasound technology
that provides a unique cross-section display of the morphology and histological
properties of the arteries. Figure 1a shows a good example of an IVUS image.
In this kind of image we apply deformable models to distinguish between fibrous
tissues and the rest. The other kind of images are color intestinal images recorded
by a capsule endoscopy. Using this kind of image, we are trying to segment the
turbid liquid. Figure 1b shows an example of this modality. We can observe in
the image that on the right side we find the turbid liquid — a liquid usually mixed
with bubbles.
The layout of the chapter is as follows: First, a mandatory analysis of the
current geometric formulations and the basics of the
STOP and GO
formulation
are provided. Second, the segmentation pipeline for deformable models is in-
troduced and three different alternative spaces are described. In this section we
also provide two tools for enhancing mask estimations. Third, the design of the
