Improving the Initialization, Convergence, and Memory Utilization for Deformable Models - Biomedical Image Analysis - page 362

scientific visualization applications, are the octrees [64, 71] and a k-d -tree-based
technique called meta-cell [15].
In [27, 28] we show that the meta-cell technique is the most suitable data
structure to perform out-of-core implementations of segmentation methods. We
take advantage of the meta-cell method to present an out-of-core implementation
of the segmentation approach proposed in [63]. This method is a straightforward
extension of the initialization method that we proposed in [26, 29].
The core of the algorithm is an out-of-core T-surfaces method based on the
meta-cell structure. To our knowledge, it is the first out-of-core algorithm for
deformable surface model reported in the literature. Besides, other parametric
deformable models as well as implicit models (level sets) and region growing
methods can be out-of-core implemented by using the same meta-cell structure
(see Section 7.10). It is important to highlight that the proposed structure is
useful not only to efficiently swap data between memory and disk, but also to
accelerate the segmentation process, as we shall demonstrate (Section 7.9).
To make this text self-contained, some background is offered in Section 7.2.
We describe the deformable model methods that will be used in this chapter.
Next, the initialization techniques of interest are described (Section 7.3).
We survey the most important works in this subject and show that their ba-
sic limitation is that the obtained contour may suffer self-intersections during
its evolution. Thus, a deformable model that can deal with such a problem is
necessary. T-snakes (or T-surfaces) is a possibility.
Thus, in Section 7.4 we describe an efficient method to initialize the T-
surfaces model, which encompasses the basic elements of the segmentation
approach presented on Section 7.5. Despite the capabilities of our segmentation
approach, we may have problems due to memory limitations for large datasets
and poor convergence for noisy images. These problems are considered in
Sections 7.6 and 7.8, respectively.
Finally, discussions and perspectives for deformable models in medical im-
ages are offered (Section 7.10). Conclusions are given in Section 7.11.
7.2 Background in Deformable Models
In some sense, deformable models used in segmentation and shape recovery
applications can be classified into two groups: free form and shape models [53].