| The cerebellum is instrumental in coordinating many vital functions ranging from speech and balance to eye movement. The effect of cerebellar pathology on these functions is typically examined using volumetric studies that depend on consistent and accurate delineation. However, no automated methods exist that can adequately delineate the cerebellar lobules, and delineation by human expert raters, the current gold standard, is time-consuming and expensive. In this thesis, three primary contributions are given that advance automated cerebellar segmentation methodology. First, a novel multiple-object extension to the level set segmentation framework (MGDM) is proposed. It is efficient, converges rapidly while avoiding local minima, preserves topology, and enables speeds to be specified on boundaries between objects rather than on the objects themselves. MGDM serves as the mechanism that optimizes an initial labeling of the cerebellar lobules while preserving topology. Second, a system was designed and evaluated that enables ground-truth data to be obtained rapidly and economically by allowing multiple inexpert human raters to delineate the cerebellum. Our system includes a hierarchical delineation protocol, a rapid rater review, and statistical label fusion steps. We also propose a topology preserving statistical label fusion that could be useful in future applications of this system. Finally, an automated method using MGDM to segment the cerebellar lobules from MR images was developed, coded, and evaluated. A topologically correct segmentation is achieved by preserving the topology of the initialization during MGDM's evolution. Speeds are derived from a probabilistic atlas, tissue segmentation, and classification of boundary voxels. We compared our method to segmentations obtained using the atlas-based segmentation and multi-atlas fusion methods and demonstrate its superior performance. |
