Elastic model-based segmentation of 2-D and 3-D neuroradiological data sets

The general problem addressed in this thesis is the automatic model-based segmentation of two-and three-dimensional objects from medical images and volumetric image data. The development of the 2-D segmentation procedure closely follows the seminal work of [Cootes et al. 1994a]. The work done being necessary to allow the extension of this technique to 3-D, presenting solutions to the difficulties arising in application to true volume data as opposed to slice-by-slice 2-D processing. The segmentation system includes both the building of statistical models and the automatic segmentation of unseen image data sets via a constrained elastic deformation of models. Geometric models are derived from a sample set of image data which has been segmented by experts; a training set. The surfaces of these binary objects are converted into a parametric surface net which is normalized to yield an invariant object-centered coordinate system. Surface descriptions are expanded into series of spherical harmonics which provide parametric representations of object shapes. Gray-level information is represented by 1-D profiles normal to the surface. The alignment is based on the stereo-tactic coordinate system, primarily chosen for its general acceptibility in applications of neuroimaging. Shape statistics are calculated from the parametric shape representations rather than from the spatial coordinates of sets of points. After initializing the mean shape in a new data set on the basis of the alignment coordinates, the model elastically deforms in accordance to displacement forces across the surface, restricted only by shape deformation constraints. The technique has been applied to the segmentation of the corpus callosum and the left hippocampus from a large series of 2-D and 3-D magnetic resonance scans. Validation was done by comparing the results of our automated segmentation approach to those performed manually by experts. 1 Acknowledgements