The principle behind these techniques is rooted in the fact that the models used in this system are approximate. The inherent ``fuzziness'' of our models was originally defined in [1] when first examining the problem space of MR brain images. Due to the variance of brain and tumor tissue between patients, and even tumor types, it is nearly impossible to create a quantitative or exact model of each brain tissue type. By allowing some ambiguity or fuzziness in the model, however, we can more easily discriminate between tissue types while still retaining accuracy. Each tissue type has defining characteristics, both in feature space and spatially. As a result, clusters suspected of belonging to a specific class can be matched against the approximate model for that class. In [1,2,3], this model matching method was crucial, both for normal/abnormal classification, as well as labeling of normal tissues. In this work, model matching is used with a slight difference. Since the primary goal of this system is to segment and label tumor, the model matching rules are used with focus-of-attention to exclude clusters not of interest (non-tumorous). Some labeling of CSF and gray matter tissues does take place in the course of isolating pixels, but the boundaries between them and non-tumorous pathology are relatively coarse.
For segmenting and labeling tumor, both rules concerning tumor in general, as well as specific rules for glioblastoma multiformes are used. Glioblastoma multiformes, the tumor types looked for in this system, are compact in nature and small in size when compared with the rest of the pathological tissues. Rules handling these characteristics complement the rules covering distribution characteristics in order to enhance segmentation quality. When other tumor types are addressed, rules for modeling and matching them will be integrated with the rule base.