Not So Elementary: Experts Debate the Takeover of Radiology by Machines

Radiologists could be replaced by computers in 20 years — or not, depending on who you were listening to during the Wednesday Controversy Session, "Elementary My Dear Watson: Will Machines Replace Radiologists?"

by Richard Dargan

Panelists John Eng, MD, Bradley J. Erickson, MD, PhD, and Eliot Siegel, MD, participated in the spirtited debate. Dr. Erickson, of the Mayo Clinic in Rochester, Minn., said that improvements in graphic processing units (GPUs) and developments like deep learning (DL) have enabled computers to surpass humans in some cases of image recognition. He cited the potential of DL to improve radiology by identifying normal screening exams and delivering high quality preliminary reports. In five years, DL will likely be able to create reports for mammography and chest x-rays, he said, and in 15 to 20 years for most of diagnostic imaging.

John Eng, MD

But co-presenter Dr. Siegel, of the University of Maryland Medical Center in Baltimore, argued that these image recognition improvements are not applicable to radiology.

"Radiology represents a completely different challenge, with much larger and more complex information," he said. "The information is extraordinarily more complex than picking out a dog or a cat. There are so many reasons why it is silly to think we'll be replaced in 20 years or in our lifetimes."

Dr. Siegel expressed concern that the hype around machine learning (ML) is becoming a major and unfounded source of anxiety among radiologists that could hurt recruitment in medical schools. He cited a story in the September 2016 Journal of the American College of Radiology that described machine learning as an "ultimate threat" that could "end radiology as a thriving specialty." Two radiology residents recently emailed him asking if they should quit the practice or risk not finding jobs when they graduate.

On the contrary, Dr. Siegel predicted that there will be more radiologists in 20 years, not fewer, and that computers will be regarded as trusted friends, able to create preliminary reports, but not primary ones.

Bradley J. Erickson, MD, PhD

The implementation of DL in radiology faces other hurdles, including the amount of time and money needed to train a machine to learn from vast databases like the National Lung Cancer Screening Trial, Dr. Siegel said. Also, the U.S. Food and Drug Administration (FDA) would be hesitant to approve technology that elevated computers to healthcare decision makers, he said, adding that medicolegal issues abound.

"Who do you sue when a computer that replaced radiologists makes a mistake, even assuming you get FDA clearance?" Dr. Siegel asked.

Dr. Erickson countered that massive investment in the DL space and its associated political power would make regulatory bodies move faster to approve new roles for computers in radiology. He also pointed to the exponentially faster computing processing power as a harbinger of a greater role for DL.

Dr. Siegel remained unimpressed, noting that processing speed is largely irrelevant if the computer is making mistakes in diagnosis.

Machines Could Make Radiology More Vibrant

Eliot Siegel, MD

Despite the good-natured ribbing, the two radiologists reached something of a consensus at the close of the session. They agreed that, in the future, computers will be performing many tasks performed by radiologists today, and that they provide a useful service in areas like quantitative imaging, biometric measures, workflow and patient safety.

"It's a natural reaction for radiologists to think the computer is going to replace them, but this fear represents an oversimplification of what a computer can do and what the profession of radiology is," Dr. Erickson said. "What machine learning can do is help remove the humdrum and make the profession more exciting and vibrant."

"Radiologists judge, explain, quality check, counsel, teach, discover, console, explore, create and dozens of other things computers aren't even close to being able to do," Dr. Siegel added.