         
|
GRASP Lab Seminar 2003-2004
October 17, 11:00 AM, Levine Hall 307, hosted by Jim Gee.
Stephen Pizer
University of North Carolina at Chapel Hill
Statistics of the Geometry of Object Populations
Abstract: Probability distributions on geometric representations of objects or multi-object complexes are useful as priors in segmentation via posterior optimization of deformable models, for characterization of geometric differences between populations, e.g., of healthy and diseased anatomy, and for a variety of other objectives. These applications require localization, accurate estimation of probabilities from a few tens of sample cases, rich and intuitive characterization of geometric effects, and null probabilities for geometrically illegal objects. I will present an implemented mechanism for achieving these aims that is based on Markov random fields of geometric residues through multiple scale levels, considering objects as geometric transformations in Lie groups, finding means and principal geodesic components on these Lie groups, and medial primitives. Probability distributions for a few anatomic objects and multi-object ensembles will be presented. While the mathematical level of this material is deep, the presentation level will be accessible to users and developers of image analysis methods.
Biography: PhD, Harvard, 1967. Dissertation work done at MGH is perhaps the first dissertation in medical image computing. On faculty in Computer Science, Radiology, Radiation Oncology, and Biomedical Engineering at UNC since 1967. Kenan Professor & advisee of about 45 dissertations. Head of MIDAG, the UNC Medical Image Display & Analysis Group, a group of over 100 from 11 departments.
