State of the Art of Level Set Methods in Segmentation and Registration of Medical Imaging Modalities - Biomedical Image Analysis - page 80

either the object or the background. The topology of an object is defined through
its number of connected components, cavities and handles. The algorithm was
implemented with a narrow-band update and 3D level set fast marching prop-
agation scheme for computational efficiency. The final object surface, which
corresponds to the zero-level of the level set function is extracted with a con-
nectivity consistent marching cubes algorithm (CCMC) [95, 96]. This algorithm
is a modification of the standard marching cubes algorithm where the resolution
of ambiguous cubes depends on predefined digital connectivity rules.
Experiments: The authors performed two sets of experiments on brain
MRIs.
1. The first experiment compared visual quality of segmented data with
the topology preserving deformable model to a standard geometric deformable
model and a parametric deformable model using the same initialization scheme.
Results showed very similar looking surfaces for the three methods but close
inspection revealed critical differences:
the parametric deformable model surface had self-intersection points,
the number of handles with the simple geometric deformable model was
40 versus 0 for the two other methods (corresponding to the correct
manifold).
In this experiment the authors also provided an example from a brain MRI
data set where part of the WM seemed to display a handle when viewed in 3D
corresponding to an incorrect topology. This type of errors, mostly due to MRI
noise, can only be corrected with a topology preserving segmentation method
such as the proposed level set framework.
2. A second set of experiments employed 21 T1-weighted MRI volumes with
voxel size (0 . 9375 × 0 . 9375 × 1 . 5 mm) from the public database of the Baltimore
study on aging [97]. Volumes were preprocessed to remove extracranial tissues,
cerebellum and brain stem. The experiments were performed with digital con-
nectivity defined as: 18-connectivity for the object (WM) and 6-connectivity for
the background. The experiments focused on the extraction of central cortical
surfaces. Prior to segmentation, the volumes were processed with fuzzy con-
nectedness [98] for labeling into memberships to different tissue types. The
result of this labeling was used as an initial segmentation of the WM and used
to fill the ventricles and the sub cortical structures (including the thalamus,