It may sound like a bird-brained idea, but scientists have trained pigeons to spot cancer in images of biopsied tissue. Individually, the avian analysts can't quite match the accuracy of professional pathologists. But as a flock, they did as well as trained humans, according to a new study appearing this week in PLOS ONE.
Cancer diagnosis often begins as a visual challenge: Does this lumpy spot in a mammogram image justify a biopsy? And do cells in biopsy slides look malignant or benign? Training doctors and medical technicians to tell the difference is expensive and time-consuming, and computers aren't yet up to the task. To see whether a different type of trainee could do better, a team led by Richard Levenson, a pathologist and technologist at the University of California, Davis, and Edward Wasserman, a psychologist at the University of Iowa, in Iowa City, turned to pigeons.
In spite of their limited intellect, the bobble-headed birds have certain advantages. They have excellent visual systems, similar to, if not better than, a human's. They sense five different colors as opposed to our three, and they don’t “fill in” the gaps like we do when expected shapes are missing.
However, training animals to do a sophisticated task is tricky. Animals can pick up on unintentional cues from their trainers and other humans that may help them correctly solve problems. For example, a famous 20th century horse named Clever Hans was purportedly able to do simple arithmetic, but was later shown to be observing the reactions of his human audience. And although animals can perform extremely well on tasks that are confined to limited circumstances, overtraining on one set of materials can lead to total inaccuracy when the same information is conveyed slightly differently.
To avoid the Clever Hans effect, researchers had 16 pigeons do all their learning once per day in a box with a computer screen without humans visible. Previously diagnosed histopathology slides from biopsies of breast tissue appeared on a computer touchscreen along with a yellow and a blue button. If the birds correctly identified cancer, they were automatically rewarded by the computer with a food pellet. If they were wrong, they got nothing. The computer not only randomized the order of images from benign or malignant tissue, but also whether yellow or blue signified "cancer" for any particular bird, to make sure the color itself didn't introduce bias. And to ensure they weren't just memorizing the slides, the birds were challenged with images of the same tissue with different magnifications and color.
Pigeons might not be ready to become certified pathologists, but a few birds in the lab might be worth a technician or two. The pigeons learned in only a matter of hours to do better than random at distinguishing cancerous from noncancerous cells. And over the course of just 1 month, their accuracy rose as high as 80% -- good, but not as good as human experts. Far more impressive was the wisdom of the flock. By showing the same images to different birds and combining their guesses, the accuracy rose to 99%, on par with trained human experts and far more reliable than a computer doing automatic image analysis.
The study is "solid and ambitious," says Aldo Badano, a physicist at the U.S. Food and Drug Administration in Silver Spring, Maryland. But he cautions that the birds are far from being useful for medicine. Just like humans, they were less accurate in diagnosing images that looked different because of changes in color and compression. And, just like humans, Badano says the pigeons might be susceptible to false-positive features. To test that, the birds must be trained on sets of images from benign tumors that appear to be malignant. He also points out that the birds probably can't be trained to take the clinical context into account in their decisions, which humans do easily. "More research is needed," he says.
If the challenges can be overcome, will cancer diagnosis in the future involve giant server farms of birds pecking away at images? "I doubt it," Levenson says. "I suspect that computers will get there first."