According to Dr. Saurabh Jha of the University of Pennsylvania, AI will not replace radiologists anytime soon. Algorithms are still not accurate enough and need human supervision.

To replace you, how good of an algorithm would have to be?

With the emergence of ChatGPT and other AI programs that can hold conversations, compose stories, and even create songs and images in seconds, this is a new question for a lot of professionals.

However, as additional algorithms promise to increase accuracy, speed up labor, and, in some circumstances, take over entire portions of the job, artificial intelligence (AI) has loomed large for clinicians who evaluate scans to discover cancer and other ailments for roughly ten years. Forecasts have ranged from gloomy futures when AI completely replaces radiologists with apocalyptic possibilities where it frees them up to concentrate on the most fulfilling aspects of their employment.

The tension is a reflection of the way AI is being implemented in healthcare. Beyond the technology itself, a lot hinges on how ready medical professionals are to entrust their patients’ health and well-being to ever-more complex algorithms that few people comprehend.

Divergent views exist within the profession over the extent to which radiologists ought to adopt technology.

“Some of the AI techniques are so good, frankly, I think we should be doing them now,” said Dr. Ronald Summers, a radiologist and AI researcher at the National Institutes of Health. “Why are we letting that information just sit on the table?”

Computer-aided imaging algorithms that can identify diabetes, osteoporosis, colon cancer, and other diseases have been developed by Summers’ lab. He credits the “culture of medicine,” among other things, for the lack of widespread adoption of any of those.

Since the 1990s, radiologists have been able to employ computers to improve images and identify areas of concern. However, the most recent AI systems are capable of far more—they can interpret the scans, provide a diagnosis, and even create written reports summarizing their findings. Millions of X-rays and other pictures gathered from clinics and hospitals are frequently used to train the algorithms.

The FDA has approved over 700 AI algorithms to support doctors in the field of medicine. Even though over 75% of them work in radiology, just 2% of radiology practices employ this kind of technology, based on a recent estimate.

Radiologists have several reasons to be cautious of AI programs despite the industry’s best efforts, including insufficient testing in real-world scenarios, a lack of openness on the algorithms’ operation, and concerns regarding the patient demographics that were used to train the programs.

“If we don’t know on what cases the AI was tested, or whether those cases are similar to the kinds of patients we see in our practice, there’s just a question in everyone’s mind as to whether these are going to work for us,” said Dr. Curtis Langlotz, a radiologist who runs an AI research center at Stanford University.

Thus far, every FDA-approved program necessitates human oversight.

The FDA convened a two-day session at the beginning of 2020 to talk about algorithms that could function without human supervision. In a letter sent to regulators shortly after, radiology experts expressed their “strong belief that it is premature for the FDA to consider approval or clearance” of such devices.

However, the first completely automated software that analyzes and generates results for chest X-rays that appear normal and healthy was allowed by European regulators in 2022. Oxipit, the app’s developer, is sending the FDA its U.S. application.

Europe has an acute need for this kind of equipment, as a lack of radiologists has left several hospitals with months-long backlogs of scans.

That kind of automated screening is probably years away in the United States. According to AI executives, radiologists are still uneasy about delegating even mundane work to algorithms, not because the technology isn’t ready.

“We try to tell them they’re overtreating people and they’re wasting a ton of time and resources,” said Chad McClennan, CEO of Koios Medical, which sells an AI tool for ultrasounds of the thyroid, the vast majority of which are not cancerous. “We tell them, ‘Let the machine look at it, you sign the report and be done with it.’”

According to McClennan, radiologists frequently overestimate their accuracy. According to research conducted by his organization, doctors who were looking at identical breast images disagreed more than 30% of the time over whether or not to perform a biopsy. After seeing the identical photos a month later, the same radiologists differed 20% of the time with their first assessments.

The National Cancer Institute estimates that during routine mammograms, 20% of breast cancers are overlooked.

Not to mention the possibility of cost reductions. The Department of Labor estimates that radiologists in the United States make over $350,000 a year on average.

According to experts, AI will soon function similarly to an aircraft’s autopilot system, carrying out crucial navigational tasks but always under a human pilot’s supervision.

That strategy provides comfort to radiologists as well as patients, according to Dr. Laurie Margolies of the Mount Sinai medical system in New York. The system obtains a second opinion on mammography ultrasounds using Koios breast imaging AI.

“I will tell patients, ‘I looked at it, and the computer looked at it, and we both agree,’” Margolies said. “Hearing me say that we both agree, I think that gives the patient an even greater level of confidence.”

The initial extensive and meticulous experiments comparing radiologists helped by AI to those working alone hint at the possible enhancements.

According to preliminary findings from a Swedish study involving 80,000 women, a lone radiologist using AI identified 20% more tumors from mammograms than two radiologists using traditional methods.

Two radiologists evaluate mammograms in Europe to increase accuracy. With just roughly 70 breast radiologists in a nation of 10 million, Sweden, like other nations, suffers from a scarcity of workers.

The study found that replacing a second reviewer with AI reduced human workload by 44%.

Nonetheless, the primary author of the paper believes that a radiologist must always reach a definitive diagnosis.

According to Dr. Kristina Lang of Lund University, “that’s going to be very negative for trust in the caregiver” if an automated system misses a cancer.

One of the difficult legal problems that hasn’t been settled is who would be responsible in these situations.

As a result, radiologists will probably keep verifying every AI decision twice to avoid being blamed for mistakes. That will probably negate a lot of the anticipated advantages, such as decreased workload and burnout.

According to Dr. Saurabh Jha of the University of Pennsylvania, radiologists could only genuinely detach themselves from the procedure with the help of an incredibly accurate and dependable algorithm.

Jha compares AI-assisted radiography to someone who offers to help you drive by continually pointing out everything on the road, at least until such systems become available.

“That’s not helpful,” Jha says. “If you want to help me drive then you take over the driving so that I can sit back and relax.”

This issue needs to be fixed for it to be a viable option. Take the example of Gemini, Google’s AI, which was recently reported by GreatGameIndia to have faced criticism for historical inaccuracies and woke biases. Google co-founder Sergey Brin admitted errors in Gemini, citing insufficient testing at the Gemini Hackathon.