Biomedical Engineering Reference
In-Depth Information
8.4 Segmentation From Multiple Nonuniform
Volume Datasets
Many of today's volumetric datasets are generated by medical MR, CT, and other
scanners. A typical 3D scan has a relatively high resolution in the scanning X -
Y plane, but much lower resolution in the axial Z direction. The difference in
resolution between the in-plane and out-of-plane samplings can easily range
between a factor of 5 and 10, see Fig. 8.9. This occurs both because of phys-
ical constraints on the thickness of the tissue to be excited during scanning
(MR), total tissue irradiation (CT), and scanning time restrictions. Even when
time is not an issue, most scanners are by design incapable of sampling with
high resolution in the out-of-plane direction, producing anisotropic “brick-like”
voxels.
The nonuniform sampling of an object or a patient can create certain prob-
lems. The inadequate resolution in the Z direction implies that small or thin
structures will not be properly sampled, making it difficult to capture them dur-
ing surface reconstruction and object segmentation. One way to address this
problem is to scan the same object from multiple directions, with the hope that
the small structures will be adequately sampled in one of the scans. Generating
several scans of the same object then raises the question of how to properly
combine the information contained in these multiple datasets. Simply merging
the individual scans does not necessarily assemble enough samples to produce
a high resolution volumetric model. To address this problem we have developed
a method for deforming a level set model using velocity information derived
from multiple volume datasets with nonuniform resolution in order to produce
a single high-resolution 3D model [43]. The method locally approximates the
values of the multiple datasets by fitting a distance-weighted polynomial using
moving least-squares (MLS) [44, 45]. Directional 3D edge information that may
be used during the surface deformation stage is readily derived from MLS, and
integrated within our segmentation framework.
The proposed method has several beneficial properties. Instead of merg-
ing all of the input volumes by global resampling (interpolation), we locally
approximate the derivatives of the intensity values by MLS. This local versus
global approach is feasible because the level set surface deformation only re-
quires edge information in a narrow band around the surface. Consequently, the
Search WWH ::




Custom Search