LANS Informal Seminar: Gordon L Kindlmann

Seminar Title: Optimizing Particle Systems for Image Feature Sampling
Speaker: Gordon L Kindlmann, Assistant Professor, Department of Computer Science, University of Chicago

Date/Time: 2010-11-10 15:00
Location: Bldg 240 Conference Center 1404-1405

Description:
Particle systems have been proposed as a means to detect and sample features in 3D image datasets, for example to find airways and lobe fissures in CT scans of the lungs, and white matter in diffusion MRI. Particles move relative to each other and throughout the image domain in order to minimize a carefully designed energy term, so that the converged configuration represents a uniform spatial sampling of the relevant image features. A significant challenge to more widespread use of particle systems in these applications is the long computation time required to run the system to convergence. Poor choices for system parameters and for the energy minimization method (stochastic search vs gradient descent) will increase the total computational cost of finding a converged configuration. Although the system will tend to convergence (in the sense of *energy* minimization) even with a wide range of system parameters, an interesting object of study is minimizing the total computational *cost* of finding the minimization. The total computational cost depends on costs of various individual operations (finding energy, finding gradients, re-applying constraints, etc) whose relative frequency depends on choices of minimization method. I will describe early efforts in modeling computational costs of particle system minimization, with the hope of applying expertise from the optimization research community to the development of faster tools for biomedical image analysis.