Machine learning and artificial intelligence (AI) are two hotly discussed topics in healthcare, but many radiologists tend to fear a future in which computers replace people.

But the fact is that there’s an overwhelming amount of misinformation and myth about how the technology will impact radiology.

Eliot Siegel, MD, chief technology officer with RadSite, addressed these issues in a Feb. 21 webinar entitled, “Machine Learning and Artificial Intelligence: Hype, Myth Reality and How It Will Revolutionize the Practice of Quality Diagnostic Imaging.”

“It’s not unusual that I’m asked by many in radiology about whether the specialty will exist in the next few years,” said Siegel, the chief of imaging services at Veterans Affairs Maryland Healthcare System.

Those questions are fueled partly by AI experts like Ezekiel Emanuel, an architect of the Affordable Care Act, who suggested radiologists may be replaced by computers in the next four to five years during a 2016 keynote at the American College of Radiology annual meeting, and Geoffrey Hinton, an engineering fellow at Google and emeritus professor at the University of Toronto, who compared radiologists to Wile E. Coyote—"you’re already over the edge of the cliff, but you haven’t yet looked down,” cited Siegel in the presentation.

However, Siegel notes his background in radiology and computer science provides him with a deeper understanding of the complexities in being a radiologist, and he envisions a brighter future.

“I believe that medical imaging is most likely to be the hardest specialty for AI software to replace within medicine, while many other specialties will be much easier to replace,” he said.

What’s wrong with this picture?

Siegel presented a cartoon image popularly found in Highlights Magazine of a farm setting designed for children to identify “What’s wrong with the picture?”

Most can quickly point to a duck watching TV or a dog holding a toothbrush. But an algorithm is not capable of doing the same. While a computer can accurately pick out images of a TV or toothbrush, it cannot diagnose an image like radiologists are able to do.

This was a key distinction Siegel highlighted throughout the presentation.

“It’s something no supercomputer or algorithm for however many billions of dollars are invested at this point—we have nothing that can beat a five or six-year-old at that task,” he said.

“The problem in radiology is that the fundamental task is much more determining what’s wrong with an image than it is recognizing objects in an image,” Siegel added.

It won’t be as easy as you hear

It’s very difficult to test humans on their competency level in radiology, and the American Board of Radiology struggles with that constantly, Siegel noted.

Even if a robot could interpret radiology studies, it would take years to test it on all anatomic areas, modalities and disease processes. It would also take to 20 or 30 years to receive FDA approval, he stated.

Similarly, Sigel argued that although we understand the data we use to teach algorithms, our inability to know how or why an algorithm is making specific predictions represents a major hurdle to gaining regulatory acceptance and an obstacle in instructing healthcare workers on how to use the technology.

He calls this the “black box” nature of machine learning.

These roadblocks are a few of many which Siegel put forth during the webinar. And although he readily admits advances in AI have progressed more quickly than 10 or 20 years ago, a long road remains before we see computers replacing radiologists.

“Radiologists do so much more than just make findings on images…dozens of things computers can’t even begin to do,” he said. “Computers will perform much more quantitative imaging and assessment, [but] radiologists are much harder to replace than has been appreciated, and I believe we’ll need more radiologists in the future and not less.”

Listen to a complete recording of the webinar here.