With the recluster mask, the first stage of reclustering can be applied. At this point, there are three primary classes of interest: CSF, gray matter, and pathology. Tumor is treated as part of the pathology class, since we would first like to remove as much ``normal'' tissue before concerning ourselves with tumor versus other pathology.
Like [1,2,3], over-clustering is used to enhance the separation between the three classes. In this case, seven clusters was found through empirical observation to be sufficient. For the purposes of strict tumor segmentation, this stage could have been performed with fewer clusters, but as one of our long range goals is complete tissue labeling, fewer clusters could not guarantee that CSF and edema (which makes up the majority of the pathological region) would not be undersegmented. Therefore, the extra clusters were used to make it easier to later label CSF, gray matter, and non-tumor pathology.
To aid this reclustering step, a primitive form of knowledge is used to initialize the cluster-center matrix in FCM. For each of the three features, T1, PD, and T2, respectively, the Group 2 cluster centers are taken and sorted from lowest to highest. This sorting give us an approximate range for the pixels being reclustered and with knowledge gathered in the previous stage (mentioned in Section 2.2) concerning cluster distribution, FCM performance can be enhanced by setting one of the seven clusters to areas in feature space were we would expect to find a specific tissue type. For example, CSF tends to have a low T1 value, a medium to high PD value, and a high T2 value. Therefore, in the cluster-center matrix, a row is seeded to reflect this region in feature space - the actual values based upon the range found in the Group 2 clusters.
While not true training data, the reliability of any training pixels that might be extracted at this point is uncertain, the cluster center locations are used to start FCM close to a good partition to guide the algorithm towards a satisfactory solution. Furthermore, such initialization significantly reduces clustering time since FCM is effectively given a head start.