Biomedical Engineering Reference
In-Depth Information
Neural networks and Hough transforms have also been applied for initializa-
tion of deformable models [14, 74].
An other possibility to address the sensitivity to initialization is the gradient
vector flow, which is a scheme based on a vector diffusion-reaction equation.
It was introduced in [77] and can be used to obtain a more efficient image force
field [78].
Deformable models can be extended to 3D, generating deformable surface
models. Besides the described problems, a new one arises when considering
these models: memory utilization.
In general, deformable surface models make use of only the data information
along the surface when evolving the model toward the object boundary [48, 49].
However, state-of-the-art implementations of these models in general do not
account for this fact and fetch the whole volume from disk at the initialization.
Such a procedure brings limitations for large size image volumes, mainly if we
consider that, in general, deformable models need not only the image intensity
but also the image gradient [42, 49].
Nowadays, image volumes with 512
3
sampling points can be acquired in
CT scanners. Besides, other scanning techniques were developed allowing
the acquisition of a huge amount of 3D color image volumes (www.nlm.nih.
gov/research/visible/visible human.html). In these cases, the data set informa-
tion (image intensity and gradient) can be too large to fit in main memory, even
if we take the usual cut policy: In a first stage, select a subvolume (a bounding
box) that contains the structure of interest, and then segment it. When the size
of the data that must be accessed is larger than the size of main memory, some
form of virtual memory is simply required, which leads to performance problems
[20].
The analysis of large data sets is a known problem in the context of scientific
visualization [15, 24, 71]. Out-of-core techniques have been developed for scalar
and vector fields visualization and new proposals are still in progress. Among
these methods, out-of-core isosurface extraction techniques are closely related
with our work, as we shall see next.
These methods partition the data set into clusters that are stored in disk
blocks, and build a data structure to index the blocks for information retrieval
(
preprocessing step
). At run-time, the data structure is read to main memory
and traverse to find out the data blocks that must be read to main memory to
perform the isosurface generation. The most commonly used data structures, for
