AMHERST, Mass. - A group of researchers at the University of Massachusetts, in collaboration with the division of neurology and the department of radiology at Baystate Medical Center, is trying to teach computers to help interpret magnetic resonance images (MRIs) to determine how well stroke patients are responding to treatment.
More than 500,000 Americans suffer strokes each year, according to A. Bernard Pleet M.D., and chairman of the division of neurology at Baystate. A stroke occurs when an artery delivering blood to the brain becomes clogged or ruptures. "When a stroke occurs, the blood supply is disrupted, and brain cells are starved of oxygen, causing tissue death in the area. The effects of a stroke may be slight or severe, temporary or permanent. They can include weakness, difficulty with speech, or loss of memory," he explained.
After a patient suffers a stroke, doctors use MRIs to determine how much brain tissue has died, how much is healthy, and how much is injured but could recover. The area where the stroke has occurred, comprising dead and injured tissue, shows up as a bright region in the image, whereas the remaining healthy tissue shows up as a pattern of various shades of gray along with some white. UMass researchers hope to train the computer to separate the image into the lesion, consisting of the dead or injured tissue, and the healthy region. In addition, they want the computer to be able to give a reliable evaluation of the volume of the lesion. Comparing images taken before, during, and after treatment will then show quantitatively how well a patient is responding to treatment. This method, when perfected, would be quicker and more sensitive than current techniques, researchers say.
"The doctor knows what these blobs and squiggles represent," said Joseph Horowitz, of the UMass department of mathematics and statistics, pointing to an MRI hanging on the wall of his office. "The problem is getting a computer to do that." Horowitz is working with fellow mathematician Donald Geman; Edward Riseman, head of the Computer Vision Lab in the University’s computer science department; computer science professor Gary Whitten; postdoctoral research associate Yasmina Chitti; and graduate student Ben Stein.
While distinguishing dark from light may sound like an easy task, it’s harder than it sounds. The images are in fact many shades of gray, with blurry patches. "It’s not all that easy even for a person," said Horowitz. "The rule of thumb is, anything that involves prior knowledge or interpretation, and that’s difficult for a person, is at least very difficult for a computer." The reason, Horowitz explained, is that any information given to a computer must be extremely precise in order for the machine to be able to use it. "The machine does not like ambiguity," said Horowitz.
Although the MRI captures a "slice" of the brain and reveals lesions, it’s important to remember that both the brain and any lesions are actually three-dimensional objects, Horowitz noted. Mathematics is integral to computer imaging for several reasons, according to Horowitz. First, in order to give a computer instructions, those instructions must be expressed in mathematical terms. Also, MRI formation relies on quantum physics, according to which protons in the body act like tiny magnets. What makes the MRI work is our ability to encode the behavior of the protons mathematically, and then create an image based on that mathematical statement. "Statistics comes into the picture because there’s an intrinsic element of randomness at the quantum level and also in the formation of the image," Horowitz added.
Eventually, researchers hope, a physician will view a computer image of the brain, and circle a region containing a suspected lesion with a computerized "pen." This would enable the machine to focus on the area of particular concern, and break it up into lesion and non-lesion regions, without creating a massive computational burden.
"It is hoped that when the technique is perfected, neurologists will be able to watch the volume of damaged tissue while treating the patient with experimental drugs, to see which can promote a decrease in the damaged area," said Pleet, who is working on the project along with Richard Hicks M.D., and chief of the division of neuroradiology at Baystate. "This method will be much quicker and more sensitive than older techniques of doing periodic neurological examinations and trying to compare them," Pleet said. "And it will be much quicker than outcome studies in which the functional status of the stroke victim can only be determined months after the event, and is then compared with other victims who have not had the treatment," he added.
"Of course we love to do beautiful mathematics, but here we want to do sound math and statistics for a problem that’s clinically important," Horowitz said. "What I really love about applied mathematics is you have to bring to bear all your knowledge on these very hard, real problems, and sometimes you’re able to accomplish something with math or statistics that you couldn’t have done otherwise."