Biomedical Engineering Reference
In-Depth Information
image shows substantially reduced noise and imaging artifacts. It also shows a
more homogeneous distribution of the chromophor compared to the individual
brains. All these properties can potentially make registration of the atlas to a
given raw image easier, and thus may aid in further improving segmentation
accuracy.
11.4.4
Multiatlas Segmentation: A Classifier Approach
We can look at an atlas combined with a coordinate mapping from a raw image
as a special type of
classifier
. The input of the classifier is a coordinate within
the domain of the raw image. The classifier output, determined internally by
transforming that coordinate and looking up the label in the atlas at the trans-
formed location, is the label that the classifier assigns to the given raw image
coordinate.
As we have briefly mentioned before, using a different atlas leads to a dif-
ferent segmentation of a given raw image. From a classifier perspective, we can
therefore say that different atlases generate different classifiers for the same
raw image. In the pattern recognition community, it has been well-known for
some time that multiple independent classifiers can be combined, and together
consistently achieve classification accuracies, which are superior to that of any
of the original classifiers [27].
Successful applications of multiple classifier systems have been reported in
recognizing handwritten numerals [30, 83] and in speech recognition [1, 65]. In
the medical image analysis field, this principle has been applied, for example,
to multi-spectral segmentation [42] and to computer-aided diagnosis of breast
lesions [17, 47].
The particular beauty of applying a multiclassifier framework to atlas-based
segmentation is that multiple independent classifiers arise naturally from the
use of multiple atlases. In fact, multiple classifiers also arise from using the
same atlas with a different non-rigid registration method. However, adding an
additional atlas is typically easier to do than designing an additional image reg-
istration algorithm. One could, however, also apply the same basic registration
algorithm with a different regularization constraint weight (see section 11.3.4),
which would also lead to slightly different segmentations.
For the demonstration in this chapter, we performed a leave-one-out study
with only one registration method, but a population of independent atlases. Each
