Biomedical Engineering Reference
In-Depth Information
method requires updating
every voxel in the volume for each iteration
, which
means that the computation time increases as a function of the volume, rather
than the surface area, of the model. Because segmentation requires only a sin-
gle model, the calculation of solutions over the entire range of isovalues is an
unnecessary computational burden.
This problem can be avoided by the use of
narrow-band
methods, which
compute solutions only in a narrow band of voxels that surround the level set of
interest [34, 35]. In a previous work [36] we described an alternative numerical
algorithm, called the sparse-field method, that computes the geometry of only
a small subset of points in the range and requires a fraction of the computation
time required by previous algorithms. We have shown two advantages to this
method. The first is a significant improvement in computation times. The second
is increased accuracy when fitting models to forcing functions that are defined
to subvoxel accuracy.
8.3 Segmentation Framework
The level set segmentation process has two major stages, initialization and level
set surface deformation, as shown in Fig. 8.2. Each stage is equally important for
generating a correct segmentation. Within our framework a variety of core oper-
ations are available in each stage. A user must “mix-and-match” these operations
in order to produce the desired result [37]. Later sections describe specialized
operations for solving specific segmentation problems that build upon and ex-
tend the framework.
Figure
8.2:
Level
set
segmentation
stages—initialization
and
surface
deformation.
