Visual nudge improves accuracy of mammogram readings
January 26, 2012
By Diana Lutz
In 2011 — to the consternation of women everywhere — a systematic review of randomized clinical trials showed that routine mammography was of little value to younger women at average or low risk of breast cancer.
The review showed, for example, that for every 50-year-old woman whose life is prolonged by mammography, dozens are treated unnecessarily — some with harmful consequences — or treated without benefit. Hundreds are told they have breast cancer when they do not.
Cindy M. Grimm, PhD, associate professor of computer science and engineering in the School of Engineering & Applied Science at Washington University in St. Louis, was not surprised by the review, a prestigious Cochrane review of the scientific evidence for a medical treatment.
“It’s not just the mammogram that’s the problem,” she says, “it’s accurately interpreting the mammogram.
“People aren’t good at it. Even expert radiologists aren’t good at it. Results vary widely from person to person, even when people have gone through the same training.”
But Grimm thought a perceptual trick she and colleagues had invented, called subtle gaze direction, might be used to improve training.
An experiment showed that a novice could be subtly guided to follow an expert’s scanpath across a mammogram and that this subtle nudging improved the novice’s accuracy.
The experimental results will be presented at the Eye Tracking Research & Application Symposium this March.
Grimm and her colleagues say the technique, should it prove durable, is widely applicable to visual search tasks. Not only might it improve the reading of mammograms and other types of medical images, such as MRIs and PET scans, but it might also be used to improve the accuracy of airport screening and learning in virtual environments.
Directing the gaze
In this painting of a Jewish Quarter by the 19th-century Dutch painter Salomon Leonardus Verveer, the people in the center of the image are brighter than they would be in a snapshot, which draws your eye toward them. The walls of the buildings focus your attention toward the middle of the image as well. WUSTL computer scientist Cindy Grimm says similar tricks can be used to help people learn difficult visual search tasks, such as scanning mammograms for tumors.
Grimm invented subtle gaze direction together with colleagues Reynold Bailey, PhD, then her graduate student, and Ann McNamara, PhD, then of Saint Louis University, a conference acquaintance.
“I had double-majored in art and computer science as an undergraduate at the University of California, Berkeley,” Grimm says. “So I was aware that artists have all sorts of tricks for guiding viewers to look at particular areas in a painting, sometimes, in the case of narrative art, in a particular sequence.
“They might make an area brighter than the background, increase the contrast or have strong edges (borders) that attract the eye.
“Movie producers do the same thing in post processing,” Grimm says. “For example, when one actor is talking and others are listening, the audience tends to watch the talker. But the producer can direct attention to a listener’s reaction instead by changing the color or brightness of that part of the image.”
Subtle gaze direction is a high-tech version of this time-honored craft. It works, says Grimm, by exploiting the difference between peripheral and central (foveal) vision.
“We use a small area in the central part of our retina called the fovea to see detail,” she says. “But foveal vision doesn’t actually cover much of our field of view.
“If you hold out your thumb, your foveal vision — the part of your surroundings you’re actually seeing in detail — covers about the same area as your thumbnail.
“We use our foveal vision to read or drive or for other detail-oriented tasks. At the same time, we are monitoring the rest of our environment with our peripheral vision, which has lower resolution but responds faster than our foveal vision.
“When our peripheral vision picks up a stimulus, our eyes move to focus our foveal vision on it so that we can see it clearly.
“During those quick eye movements, called saccades, vision is suppressed, or masked, so that the motion of the eye, the motion blur of the image and the gap in visual perception are not noticeable to the viewer. We lose an astonishing 40 minutes of vision a day to saccadic masking.”
Grimm, Bailey et al.
In subtle gaze direction, the modulation of the brightness (middle column) or warmth (third column) of a part of the image in the peripheral field of view is used to attract the viewer’s focus to that area. By moving the stimulus the viewer can be coaxed into scanning the image in a particular pattern. The stimulus is cut off before the viewer can focus on it and so the gaze direction remains subtle.
To direct the gaze, Grimm and her colleagues changed the brightness or “warmth” of an area in the peripheral field of view to draw the novice’s focus to this area.
The stimulus remained subtle, however, because the viewer’s gaze is monitored in real-time by an eye-tracking device and the modulations to the peripheral vision are terminated before the eye fixates on them.
“The idea,” says Grimm “is to get someone to look in a particular direction while altering their experience of viewing the image as little as possible.”
“In the case of mammograms,” for example, “you want to get a learner to look at the tumor region but you don’t want to do anything that makes the tumor region look different than it does on the mammogram itself.”
The mammography study
Reading mammograms is a good target for computer assistance because training is time-consuming and expensive, typically requiring a four-year residency and a two-year fellowship.
Despite advances in technology, novices are still trained by working as an apprentice to an expert.
The mammography study, led by Bailey, now an assistant professor of computer science at the Rochester Institute of Technology, brought together the same group of scientists as the subtle gaze direction experiment. McNamara is now assistant professor of visualization at Texas A&M University.
For the study, Grimm and her colleagues used a database of images provided by the Mammographic Image Analysis Society that includes both images and text files that contains coordinates of abnormalities and their size.
“Expert diagnostic radiologists have a particular search pattern that is not the same as that of a novice,” Grimm says. “We don’t know exactly what they’re doing, but they tend to do a fairly broad scan and then fixate on parts of the image that have a tumor-like texture. A novice might instead attend to brighter spots in the image or fail to scan all of it.”
Bailey hired an expert radiologist at the Rochester Institute of Technology to view and mark 65 images from the database. The expert’s scanpath was recorded during this process by an eye-tracking system.
During the experiment, subtle gaze direction was used to guide a group of novices along the expert scanpath. A control group viewed the mammograms without gaze manipulation.
In the study, gaze direction was used to nudge novices into following an expert radiologist’s scanpath (a simplified version of which is shown in green) as they looked at a mammogram. A potential tumor is circled in red.
Novices who were guided were significantly more accurate than the control group or a third group guided along a random path. Moreover, even though the training session was brief, the effect lingered even after gaze manipulation was disabled.
To watch a Power Point of the experimental results, click here.
Grimm says more work must be done to show that more extensive training will stick long-term. In the meantime, she can think of many ways gaze manipulation could be used to improve performance on visual search tasks.
“One simple use of the technology would be to make sure readers look at every part of the image. If you’re using eye tracking,” she says, “you know where people are looking, so you can make sure they don’t skip part of the image.”
Gaze manipulation might also be used to assist tumor-recognition software. “Suppose you had a software program that was reasonably good at spotting possible tumor areas but, erring on the side of caution, flagged too many areas as suspicious.
“Such software might be paired with gaze direction to ensure the radiologist looked at all of the flagged areas,” she says. “That wouldn’t necessarily be a training application; it could be a routine element of reading mammograms.”
The mammogram study is widely applicable, Grimm says, because there are so many visual search tasks. She mentions airport scanners, but they are just at the top of a long list.
“I work with someone who identifies pollen species,” she says. “Apparently, it takes a novice a year to learn, and they spend hours and hours looking through a microscope at these pollen grains. Again, some people are good at it and others struggle for competence.
“Perhaps in that case, as well, gaze direction could be used to train novice pollen identifiers.”