There can be a significant degree of variation in tissue distributions between patients and the fuzziness in the distribution knowledge can cause ambiguity in cluster labels. Therefore it is important to have knowledge that is independent of fuzzy clustering to make the knowledge base more robust. The anatomical structure of the brain provides such a source.
In our previous efforts, anatomical knowledge was used to make normal/abnormal classifications through a ``default reasoning'' method that looked for significant deformations from our approximate models, as well as labeling clusters that passed these tests. While all slices used here were first classified as abnormal, the use of anatomical knowledge to quickly isolate the pathological tissues from normal tissues and tumor from non-tumor pathology remains important.
The knowledge extracted from anatomical structures was employed in a ``fuzzy'' fashion, allowing for greater flexibility for significantly different brain shapes found between patients. For example, CSF in normal slices is symmetrical along the vertical axis. We require only that there be approximately the same number of CSF pixels in each brain hemisphere.