Biomedical Engineering Reference
In-Depth Information
Among the four criteria that do not depend on the a posteriori accuracy
evaluation and thus a gold standard, the NMI image similarity after non-rigid
registration performed slightly better than the other three. It was therefore se-
lected as the criterion used for the SIM atlas selection strategy in the comparison
to the other three strategies later in this chapter (section 11.5.4).
11.4.3
Segmentation with an Average Shape Atlas
As we stated in the previous chapter, atlas-based segmentation is an easier task if
the atlas is similar to the image that is to be segmented. Smaller magnitudes of the
deformation between image and atlas that the non-rigid registration algorithm
has to determine typically result in a higher accuracy of the matching. If the atlas
is itself derived from an individual subject, then the risk is high that this individual
is an outlier in the population. In such a case segmenting other subjects using
the atlas becomes a more difficult problem. A better atlas would be one that is
as similar to as many individuals as possible. Such an atlas can be generated by
creating an average over many individuals.
For the human brain, such an average atlas is available from the Montreal
Neurological Institute (MNI) as the BrainWeb phantom [5, 13]. Note, however,
that the BrainWeb phantom is an atlas of brain
tissue types
, so as we discussed
in section 11.1, it is not as useful for atlas-based segmentation as an atlas of
brain structures. For the human heart, an average atlas derived from cardiac
MR images [45] has been used for atlas-based segmentation [35]. Similarly, an
average atlas of the lung has been derived from CT images [33].
For demonstration in this chapter, we have therefore generated an average
shape atlas of the structures of the bee brain using a technique outlined be-
low [51].
11.4.3.1
Iterative Shape Averaging
One way of obtaining an average shape atlas from a population of subjects is to
generate an active shape model (ASM). In short, an ASM provides a statistical
description of a population of subjects by means of an average shape and the
principal modes of variation [14, 34]. Generating an ASM typically requires the
identification of corresponding landmarks on all individuals, a tedious and error-
prone process despite recent success in automating this step using non-rigid
