“Radiology is leading the way in the development and implementation of AI in clinical practice,” he adds. But AI isn’t reducing the need for human input.

“AI is not a better kind of intelligence, it’s just a different kind of intelligence,” Langlotz says. “A human plus a machine is better than either one alone. I would say that has been true since I began studying AI in the 1980s, and it continues to be true today.”

About two-thirds of radiology departments in the United States use AI in some capacity, according to a recent unpublished survey from the American College of Radiology (ACR) Data Science Institute. The number has roughly doubled since 2019, says Christoph Wald, vice chair of the ACR’s board of chancellors and chair of its informatics commission.

Wald, who also works as a senior associate consultant radiologist at the Mayo Clinic, said that there are about 340 FDA-approved AI radiology tools to date and that the number keeps rising. The majority of these tools, Langlotz says, are detection algorithms. These can look for everything from brain tumors and pneumonia to breast cancer and strokes.

A CT scan of the body includes hundreds of images that radiologists must study. AI tools can filter through these images to figure out which ones are most likely to have abnormalities. A study published in the academic journal Neuroradiology found AI tools can effectively alert radiologists to critical findings in head CT scans (such as hemorrhage and hydrocephalus) so they can prioritize those cases.

“That allows us to bring those to the top of the list and interpret them quicker,” says Langlotz, who also runs the Center for Artificial Intelligence in Medicine and Imaging at Stanford University. “That can have a positive effect on urgent situations, like patients in the emergency department or intensive care unit who can then have their problem addressed sooner because their images get interpreted more quickly.”

AI also has the potential to give patients more accurate results.

“Let’s say I’m an expert, the best in the world. The AI program may help me a little bit, but not a lot,” says Elliot Fishman, a radiologist at Johns Hopkins Medicine who uses AI technology for early pancreatic cancer detection. “But if I’m the average radiologist — and most people are average — when you use AI, you become an expert. Who benefits from that? The patients.”

Research has shown that when two radiologists read the same study, there is a 3 percent to 5 percent discrepancy in their findings.

Pranav Rajpurkar, an assistant professor of biomedical informatics at Harvard Medical School and the co-founder of a company called a2z Radiology AI, hopes AI will offer an “extra layer of security” by giving doctors a second read on everything.

A randomized, controlled, population-based 2023 study published in the journal Lancet Oncology points to the promise of AI in radiology. More than 80,000 women in Sweden were randomly assigned either two radiologists to read their mammogram or one radiologist plus AI. The study determined that there was a similar cancer detection rate for both groups.

Another 2023 study, in the journal Radiology, found that one AI tool was very effective at ruling out abnormalities on chest X-rays, with a sensitivity of 99.1 percent. And one 2022 study published in the journal Frontiers in Public Health found that AI was effective at detecting lung nodules on CT scans.

But AI tools aren’t perfect. A 2024 study published in Radiology looked at AI’s ability to exclude certain diseases on chest X-rays. Although AI had a high level of accuracy for excluding disease, when it made a mistake, it was potentially more critical or clinically significant than something missed by a radiologist.

Most AI detection products produce false positives that radiologists are responsible for following up on. “AI that’s focused on detecting abnormalities can actually create more work for the radiologist,” Langlotz says, adding that he believes this has slowed adoption of some of AI algorithms.

Many upcoming AI tools focus on helping radiologists with the time-consuming task of drafting reports, which could yield “significant time savings” and help in drafting “higher-quality, more consistent reports,” Langlotz says.

But the use of AI in radiology also opens up potentially complex legal questions. Currently in the United States, liability rests with the radiologist, not the AI technology company, because AI’s findings must be approved by a licensed physician, Wald says.

This could change if autonomous AI without physician oversight enters the U.S. market. An autonomous radiology tool for reading chest X-rays from a company called Oxipit has been cleared for use in the European Union. Plus, some U.S. companies are experimenting with the concept.

Although AI is a hot topic, it has been part of radiology for decades, says Despina Kontos, a computer scientist and professor of radiological sciences at Columbia University Vagelos College of Physicians and Surgeons. It used to be called computer-aided diagnosis, but it was essentially doing the same thing. “Radiology has been a bit ahead of the curve compared to other medical disciplines in the use of computers and AI,” she says.

Most experts agree that the most likely reality isn’t that radiologists will be replaced by AI, but rather that they will use the technology in different ways to improve their accuracy and eventually speed up their workflow.

“I think what you’re seeing are really wonderful, innovative, changing times,” Fishman says. “I’m not saying AI is going to replace radiologists, but AI is going to be a vital part of assisting radiologists in reading studies for a long time.”