by
Lisa Chamoff, Contributing Reporter | September 19, 2018
Dr. Eliot Siegel speaking at NYMIIS
While the fear that artificial intelligence will replace radiologists remains, Dr. Eliot Siegel, professor of diagnostic radiology at the University of Maryland School of Medicine, reassured the physicians and technicians gathered Monday for the New York Medical Imaging Informatics Symposium (NYMIIS) that their jobs are safe in the near future by taking a page from a children’s magazine.
During his talk, Machine Learning and Artificial Intelligence: Hype, Myth and Reality and How It Will Revolutionize the Practice of Diagnostic Imaging, Siegel showed an image from the children’s magazine Highlights, where readers were supposed to point out what was wrong with the picture.
The image included a television on the roof of a house, a snowman on a sunny day and a farmer holding a giant fork and spoon.
“What the computer could do really, really well is tell you that there’s a TV, and a fork and a spoon, and it could tell you that there is a snowman, but what it couldn’t do is tell you what’s wrong with this picture, and that’s a very important distinction,” Siegel said. “I think a lot of folks have confused what’s in this picture with what’s wrong with this picture. There’s no computer system at any level of sophistication that can beat a five-year-old on this Highlights magazine challenge. That’s part of the reason why radiology is so hard and so different from the things where there have been success.”
In order to replace radiologists, there will need to be what Siegel called “general artificial intelligence,” which includes seeing health issues in a larger context.
“In order to replace a radiologist, I think we’re going to have to achieve something that is much more complex, approaching artificial general intelligence, where a computer could reason and plan and solve problems, think abstractly,” Siegel said. “Creating artificial general intelligence is a much harder task and we’re nowhere near close to it.”
Machine learning will be useful for radiology because of the ability to recognize trends from data to make proper predictions.
Siegel gave an example of using data from the National Lung Screening Trial to predict cancer rates. One could assume that big lung nodules are more likely to be cancerous than small ones, and that irregular ones are more likely to be cancer than smooth ones, but larger nodules start becoming smoother, he explained.
