Seminars/Colloquia/Invited Talks

Seminars

Erik G. Learned-Miller

The Power of Multi-Image Methods in Medical Imaging: Joint Non-Parametric Statistical Models for Registration and Bias Removal

PLACE:	Clark 314
EVENT:	CIS Seminar Series
DATE:	February 21, 2006
TIME:	1:00 - 2:00

Abstract

In this talk, I will present a family of methods called "congealing" for the modeling of sets of images, volumes, or other data arrays over continuous spaces. The simplest form of congealing is joint alignment, where a set of images are simultaneously warped to be as similar as possible to each other. Our "joint similarity" criterion is defined in terms of maximizing the likelihood of the pixels (or voxels) in a set of images under a non-parametric model defined by the other images. A set of images is "evolved" to maximize this criterion until a stable point is reached, at which point the images are as much like each other as possible under a given family of transformations. This process results in the joint alignment of a set of medical volumes without having to start with a single reference volume or model.
In addition to using congealing for joint alignment, we show how the same method can be used for the problem of removing the so-called "bias" from magnetic resonance images. Once again, under a particular set of transformations, the images are driven toward each other, until their "self-likelihood" is as high as possible. It turns out that this process is equivalent to minimizing the pixelwise (or voxelwise) entropies of the images. We argue that these methods are ultimately more meaningful than methods based upon discrete tissue models and EM-like procedures for alignment or bias removal.

Brief Biography

Erik G. Learned-Miller is an Assistant Professor in the Department of Computer Science at University of Massachusetts Amherst. His interests can be broadly categorized as applying ideas and methods from machine learning to problems in machine vision. Problems he has worked on include learning from a small number of examples, independent component analysis, learned color constancy, developing probability models of shape deformation, and mathematical expression recognition. In his spare time, he is the volunteer Production Editor for the Journal of Machine Learning Research (JMLR).