Automatic segmentation of MR volumes of the human brain is a complex task. This paper presents an approach that combines knowledge based techniques with unsupervised fuzzy clustering to completely segment and label glioblastoma multiforme tumors. Each slice has been previously over-segmented into ten classes, found to contain tumor, and had it's extra-cranial and white matter tissues identified. This information was then provided to this system.
Like its predecessor systems, over-clustering plays an important role here. By providing more clusters than there are tissue types, the amount of under-segmentation was greatly reduced, allowing objects of interest to become more easily identifiable. At the start of this system, pixels not identified during pre-processing are over-clustered and a search is done for the normal tissues, gray matter and CSF. When these were identified and removed, only pathological tissues remained. These tissues were then reclustered, also using over-clustering, to separate glioblastoma multiforme tumor from other pathology such as edema and necrosis. The glioblastoma multiforme was then segmented and labeled. At each stage, knowledge was crucial in proving both information concerning tissue distribution in feature space, as well as spatial characteristics.
The advantages of the proposed approach are demonstrated by the successful performance of the system on the thirteen slices. The final segmentation of glioblastoma multiforme tumor compares favorably with hand-labeled ``ground truth'' images of the tumor. Some partial tissue labeling is achieved, their final borders were not of concern at this point. Rules could be later developed to fully label other tissues.
Acknowledgements: This research was partially supported by a grant from the Whitaker foundation and a grant from the National Cancer Institute (CA59 425-01).