Nine years ago, one of the world's leading artificial intelligence scientists selected an endangered occupation.
“People should now stop training radiologists,” Jeffrey Hinton said, adding that it is “completely obvious” that AI will outperform people in that field within five years.
Today, radiologists – medical imaging physicians looking inside the body to diagnose and treat illnesses – remain in high demand. A recent study from the US University of Radiation predicted a steady growth workforce until 2055.
Dr. Hinton, who won the Nobel Prize in physics for his pioneering research in AI last year, was very right to have a significant impact on technology.
This applies to radiologists at Mayo Clinic, one of the nation's leading healthcare systems, whose main campus is in Rochester, Minnesota. In recent years, they have begun using AI to scrape images, automate daily tasks, identify medical abnormalities, and predict illnesses. AI also functions as a “second eye.”
“But will that replace radiologists? We didn't think so,” said Dr. Matthew Colestrom, chairman of radiology at Mayo Clinic. “We knew how difficult it was and how it all relates.”
Computer scientists, workers experts and policy makers have long debated how AI ultimately unfolds in the workforce. Will it be a smart helper, an improved human performance, a robot agent, and drive away millions of workers?
The debate is escalating as the cutting edge technology behind the chatbot appears to be improving faster than expected. Leaders of Silicon Valley's Openai, humanity and other companies predict that AI will eat humans on most cognitive tasks within a few years. However, many researchers have predicted more gradual changes, such as electricity and the Internet, in line with past earthquake inventions.
The predicted extinction of radiologists provides a case study. So far, AI has proven to be a powerful medical tool for increasing efficiency and expanding human capabilities, rather than anyone else's work.
When it comes to the development and deployment of AI in medicine, radiology has been a major target. Of the more than 1,000 AI applications approved by the Food and Drug Administration for medical use, approximately three-quarters are radiology. AI is usually excellent at identifying and measuring certain abnormalities such as lung lesions and breast masses.
“There have been some amazing advancements, but these AI tools are looking for most things,” said Dr. Charles E. Kern Jr., a professor of radiology at the University of Pennsylvania Perelman University School of Medicine and editor of the journal Radiology: Artificial Intelligence.
Radiologists do much more than research images. They advise other doctors and surgeons, talk to patients, write reports and analyze medical records. After identifying suspected tissue clusters in organs, they interpret what it means for individual patients with a specific medical history, depriving them of years of experience.
David Ortl, a labor economist at the Massachusetts Institute of Technology, said AI “underestimates the complexity of the work people actually do.”
At Mayo Clinic, AI tools are researched, developed and tailored to fit busy physician work routines. The staff has grown by 55% to over 400 radiologists since Dr. Hinton's fate prediction.
Spurred by warnings and advances in AI fuel image recognition in 2016, radiology leaders gathered the group to assess the potential impact of the technology.
“We thought the first thing we should do is use this technology to improve us,” recalls Dr. Callstrom. “That was our first goal.”
They decided to invest. Today, the Department of Radiology has a 40 AI team, including AI scientists, radiation researchers, data analysts and software engineers. They have developed a range of AI tools, ranging from tissue analysis equipment to disease predictors.
The team works with experts like Dr. Theodora Pototzke, whose focus is on the kidneys, bladder and reproductive organs. She describes the role of a radiologist as a “other physician,” and clearly communicates imaging results, providing support and advice.
Dr. Potretzke works with AI tools to measure kidney volume. Kidney growth, when combined with cysts, can predict decline in kidney function before it appears on blood tests. In the past, she measured the volume of her kidneys primarily by hand, equivalent to an on-screen ruler. The results were varied and the chores took time.
Dr. Potretzke worked with the department's AI team as a consultant, end user and tester. She helped design software programs with color codes for various organizations and reviewed measurements.
Today, she brings up the image on a computer screen, clicks on the icon, and the kidney volume measurements are displayed immediately. She saves 15-30 minutes each time she examines the image of her kidneys and is consistently accurate.
“This is a great example of passing it very comfortably to AI for efficiency and accuracy,” Dr. Potretzke said. “It can be augmented, supported, quantified, but I'm not in a place to give up interpretive conclusions about technology.”
Under the hall, Staff radiologist Dr. Francis Buffer explained the various ways in which AI was applied to the field, often in the background. He said manufacturers of MRI and CT scanners use AI algorithms to speed up the shooting of images and clean them.
AI also automatically identifies images that show the highest probability of abnormal growth, essentially telling the radiologist “look here first.” Another program scans images of heart or lung clots, even when the medical focus is elsewhere.
“AI is now everywhere in our workflow,” Dr. Buffer said.
Overall, Mayo Clinic uses over 250 AI models, both developed internally and licensed by suppliers. The Radiology and Heart Disease Division are the largest consumers.
In some cases, new technology opens the door to insights beyond human capabilities. One AI model analyzes data from the ECG and predicts patients who are more likely to develop cardiac fibrillation, cardiac ribrillation.
Radiology research projects employ AI algorithms to identify subtle changes in pancreatic shape and texture, and detect cancer up to two years before traditional diagnosis. The Mayo Clinic team is working with other healthcare organizations to further test algorithms for more data.
“Mathematics allows us to see what the human eye can't,” said Dr. John Haramka, president of the Mayo Clinic Platform, which oversees the digital initiatives of the health system.
Dr. Halamka, an AI optimist, believes that the technology will transform medicine.
“Five years from now, not using AI will be a medical malpractice,” he said. “But that's going to mean that humans and AI will work together.”
Dr. Hinton agrees. Looking back, he believes he spoke too widely in 2016, he said in an email. He didn't make it clear that he was purely talking about image analysis and was wrong about timing, but not in direction, he added.
Over the years, most medical image interpretations are “made through a combination of AI and radiologists, which in addition to improving accuracy, radiologists become much more efficient,” Dr. Hinton said.
