Title: Statistical Characterization and Discrimination of the Geometry of Medical Objects
Speaker: Stephen M. Pizer, Kenan Professor of Computer Science, Radiology, Radiation Oncology, & Biomedical Engineering, University of North Carolina, Chapel Hill.
Co-author: Guido Gerig
Date: Tuesday, November 6, 2001
Time: 3:30-5:30 pm
Place: N104 LC (the Lindquist Center, next to the engineering building)
Statistical characterization of geometry within patient classes can be described in terms of deformations involving homologous points. Multiscale representations provide advantages in the trainability of the probability distributions to be used for characterization and discrimination of classes, and they provide locality within the statistical framework. I will explain how each of m-reps (sampled medial representations), spherical harmonic boundary representations, boundary point distribution models, and voxels grids (sampled volumes) provides a means of representing the object and multi-object geometry in this framework. I will compare the ways in which each provide the positional correspondences that identify and represent the homologies needed in statistical characterization and the strengths and weaknesses of each in providing locality. I will present results of a variety of studies in the neuroscience of brain structures that already have benefited from these approaches, as well as a Monte Carlo approach to the generation of test images for validation of segmentation. I will foreshadow results in the probabilistic geometric analysis of multiobject complexes. segmentation differences will be given.