"We are going to explore whether evolutionary computer programs can learn how to very accurately pinpoint on a CT (computed tomography) scan abnormalities in the prostate," said Mitchell. Relying on such an imaging technique to test the prostate would be far less invasive to men with high PSA levels, noted Mitchell.
About 190,000 new cases of prostate cancer are diagnosed in the United States every year, and about 30,000 men died from the disease in 2002, according to the American Cancer Society.
Evolutionary computing mimics the idea behind biological evolution. In a so-called "genetic algorithm," a population of computer programs evolve over time. Each computer program is an "organism," and each has a fitness value which measures how well it performs tasks. When computer programs have "offspring," the children are copies of their parents, though with mutations, and the process keeps repeating itself generation after generation.
The idea is that after many computations, the genetic algorithm will by natural selection produce a computer program that solves a given problem, rather than the programmer having to design it.
"You evolve the computer program, rather than build it," explained Mitchell. "This is what we'll try to do to design more advanced image processing programs."
For their study, Mitchell and Song will start with an initial population of random image processing programs. Each program is executed on an image and produces an analysis of that image by highlighting certain pixels (points in the image). Its fitness corresponds to how many pixels are highlighted correctly, compared with a "training" image that is highlighted correctly.
"When the program is finished running, you ask, How close is this to what I asked it to do? How many pixels did it get right?" said Mitchell. "Most of the offspring do horribly, but there are ones that do bette
'"/>
Contact: Sydney Clevenger
clevenge@ohsu.edu
503-748-1546
Oregon Health & Science University
9-Mar-2004