Data Driven Image Models through Continuous Joint Alignment and "Have you got acromegaly?"
Speaker: Erik Learned-Miller , UMass - Amherst
Date: April 27 2005
Time: 2:45PM to 3:45PM
Location: 32-G449 (Patil/Kiva)
Host: Greg Shakhnarovich, CSAIL
Contact: Greg Shakhnarovich, xx3-8170, gregory@csail
Relevant URL: ABSTRACT
In the first part of this talk, I will start by reviewing my "congealing"
algorithm, a method of joint alignment of images. In previous work, I used
congealing to form factored models of handwritten characters, and used these
factored models to develop a handwritten digit classifier from a single
example of each digit. After briefly reviewing this work, I will discuss new
work in which I extend the notion of "alignment" to non-spatial
transformations. In particular, I show how, with only very minor
modifications, the congealing algorithm can be used to eliminate unwanted
low-frequency components, or "bias fields" from a set of magnetic resonance
(MR) images, giving state-of-the-art results for the bias removal problem.
In the second part of the talk, I discuss the (seemingly unrelated) problem
of identifying a disease known as acromegaly from normal photographs of
patients. Acromegaly is a condition in which excess growth hormone causes
swelling of the face and new disfiguring growth of the bones in the head,
hands, and feet. If detected early enough, it can often be successfully
treated. Because most laypeople and physicians cannot recognize the disease,
it often progresses further than it would if it were detected early. I
describe a system we have built, using the 3-D, laser scan-based,
statistical, morphable, head model of Blanz and Vetter, to screen for
acromegaly from regular frontal photographs of people. Our aim is to build a
system which could be used in a locale such as the Department of Motor
Vehicles, on a purely voluntary basis, to do automatic screening for a
variety of conditions. At the user's option, a report with posterior
probabilities of each relevant disease could be printed, with
recommendations about seeing a physician.
Finally, I discuss my current goal of using congealing-like algorithms to
build statistical models such as the morphable head model, without the use
of cumbersome and expensive 3-D laser scanners.
See other events that are part of MIT Machine Vision Colloquium 2004/2005
See other events happening in April 2005
