Biomedical Engineering Reference
In-Depth Information
Chapter 11
Quo Vadis
, Atlas-Based Segmentation?
Torsten Rohlfing,
1
Robert Brandt,
2
Randolf Menzel,
3
Daniel B. Russakoff,
4
and Calvin R. Maurer, Jr.
5
11.1
Segmentation Concepts
There are many ways to segment an image, that is, to assign a semantic label
to each of its pixels or voxels. Different segmentation techniques use different
types of image information, prior knowledge about the problem at hand, and
internal constraints of the segmented geometry. Which method is the most suit-
able in any given case depends on the image data, the objects imaged, and the
type of desired output information.
Purely intensity-based classification methods [29, 76, 81] work locally, typi-
cally one voxel at a time, by clustering the space of voxel values (i.e., image inten-
sities). The clusters are often determined by an unsupervised learning method,
for example,
k
-means clustering, or derived from example segmentations [43].
Each cluster is identified with a label, and each voxel is assigned the label of the
cluster corresponding to its value. This assignment is independent of the voxel's
spatial location. Clustering methods obviously require that the label for each
voxel is determined by its value. Extensions of clustering methods that avoid
overlapping clusters work on vector-valued data, where each voxel carries a
vector of intensity values. Such data is routinely generated by multispectral
