After initial reclustering, the process of separating normal tissues from pathological tissues is begun. The majority of knowledge used here comes from the distribution of cluster centers in feature space. Spatial properties, however, also play an important role in clusters whose identity is not certain.
The first of the seven clusters to be identified is the cluster with the lowest T1 value. This cluster can be immediately labeled as CSF. In some cases, where the brain is severely deformed due to pathology, some edematous CSF (CSF with some edematous properties and possibly some edema mixed into the CSF) can be placed into this cluster. While this presents the problem of some undersegmentation between normal CSF and abnormal CSF, this does not involve tumor itself and can be ignored until we address the issue of complete tissue labeling. For now, labeling the entire lowest T1 cluster as CSF is sufficient.
With the lowest T1 cluster labeled, the two lowest PD clusters can be located and labeled as gray matter. This, like the CSF cluster labeling, was based on empirical observation of the seven cluster distribution.
Once definite CSF and gray matter clusters have been labeled and removed from further consideration, the next cluster to be located is the one closest in PD and T2 space to the second lowest PD gray matter cluster, the ``second gray'' cluster. This cluster, the ``third gray'' cluster, is located by finding the Euclidean distance in PD and T2 space of all unlabeled clusters from the second gray cluster. Once located, the third gray cluster may either be pure gray matter, pure edema, or a possible mixture of gray matter and pathology and/or CSF. For the purposes of tumor segmentation, the system is concerned with determining whether significant pathology is present. Should the cluster be just gray matter with some possible ``peripheral'' CSF (CSF lying around the brain and within folds in the hemisphere, as opposed to CSF within the ventricular area), this cluster will be masked out and removed. Otherwise, the cluster will remain for possible inclusion in the ``pathology mask,'' which will eventually hold nearly all pathology pixels, including tumor.
The makeup of the third gray cluster is determined a density
measure. The density measure is performed by first isolating
the cluster, providing a sub-image to perform a 
erosion
operation upon. The number of pixels in the image after
erosion, divided by the original number of pixels in the sub-image
will determine how dense, or compact, the cluster was spatially.
A very compact object will have far fewer pixels eroded
than that of a cluster that is dispersed throughout the image.
In such a case, edema will have a higher density value than
that of gray matter. Clusters with a significant density
value do not necessarily contain pathology, the presence of some
CSF within the cluster may increase the density value,
but the rule is designed to remove clusters with a very high
certainty of being gray matter.
The rule used by this system will mask
the cluster to gray matter, if its density value is less
than 
.