Generally, both pattern distribution and domain specific knowledge is used in guiding focus-of-attention decisions. When the system focuses on a certain spatial area, our knowledge base allows the system to either deduce or estimate the makeup of the region. This includes the number of classes (tissues) in the focussed region, and allows for the possibility of choosing ``prototypical'' patterns.
Essentially, prototypical patterns are patterns that have very high membership within a specific class. In this system, prototypical patterns are used to train a semi-supervised fuzzy clustering [7] based on the knowledge gathered from the cluster distribution in feature space, and as the focus narrows, the effectiveness of these prototypical patterns becomes much greater. Patterns from clusters known not to be of interest can also be used to help ``enhance'' existing differences between tissue types. Both types of pixels are used in Sections 4.4 and 4.5 in an enhanced form of FCM (``semi-supervised FCM'') [7] to separate pathology from normal tissue, then tumor from non-tumor pathology.