Our solution to the problem of accurately segmenting a difficult feature space was ``over-clustering,'' meaning that we deliberately clustered the pixels into a greater number of classes than the number of expected classes in the data set. This inherently leads to some level of over-segmentation, but it also reduces both the chance and frequency that different tissues are clustered into one class. This step is necessary because different tissues may be very close in some features and tend to be grouped into a single class.
Over-segmented clusters, since they are homogeneous, can be more easily identified by class-unique characteristics than can clusters that contain multiple class types. Furthermore, merging over-segmented clusters of the same tissue type is a much simpler task than splitting up under-segmented clusters. There are limits to over-clustering, however. Over-clustering does not always prevent under-segmentation, and an increase in the number of classes used for clustering will dramatically increase the amount of computation time. It is important to weigh these factors when choosing the initial number of clusters.
The primary source of knowledge for merging over-segmentations comes from the cluster distribution in feature space. In [1,2,3], this knowledge allowed us to identify under-segmented classes and merge their component clusters into one class. In this work, tumor is the class of real interest; to achieve consistent and satisfactory labeling, over-segmented tumor must be detected and merged.