What will it take to really bring AI into the DI department?

Scores of North American radiology departments – especially at academic centres – are experimenting with artificial intelligence and many have set up pilot projects. The impact on healthcare delivery and workflow, however, has been modest.

Still, many in the healthcare sector believe AI is capable of revolutionizing the delivery of care.

So, what would it take to really start using AI, with extensive usage throughout organizations?

Put another way, what would be the “killer app” that makes everyone want to have AI in their diagnostic imaging department?

This topic was discussed by a panel of DI experts during the recent RSNA conference. Sponsored by Canon Medical, it was hosted by Dr. Eliot Siegel, professor and vice-chair of radiology at the University of Maryland.

Participants included Dr. Peter Chang, a radiologist, computer scientist and entrepreneur at the University of California Irvine School of Medicine; Dr. Patrik Rogalla, professor of radiology at the University of Toronto and site director at the Toronto General Hospital; Cindy Siegel, corporate director of imaging operations at UHS, Philadelphia, Penn; and Tom Szostak, director of healthcare economics at Canon Medical.

For Dr. Peter Chang, the killer app that would spark widespread adoption of AI would be an algorithm that could tell the difference between normal and abnormal exams. That would save radiologists an incredible amount of time and allow them to focus on patients with actual problems and issues.

“If this is done, it would have a real impact on workflow,” said Dr. Chang. “It would also demonstrate to the average hospital the value of implementing AI, whereas with other algorithms, the benefits may be hard to show and adoption becomes much slower.”

Dr. Eliot Siegel agreed: “If mammo AI is able to identify even 80 percent of exams with 97 percent confidence, it would be valuable.” He noted that a level of accuracy of 97 percent “is better than most radiologists.”

However, he pointed out some complicating factors.

To get to a 97 percent accuracy rate may take some time. Training today’s AI systems – namely Deep Learning solutions that learn from their mistakes – requires huge datasets that are not yet available.

In particular, these datasets must account for all kinds of possibilities and variables.

“There are different populations, and there are patients with rare diseases who are not fairly included in many of these datasets,” asserted Dr. Siegel. “This becomes a challenge when creating algorithms.”

Dr. Patrik Rogalla concurred that “sorting out disease from no disease would be the best application you could have.”

He commented that AI will not replace radiologists – at least not in the near- or medium term, as demand for imaging has been soaring.

However, he said that radiologists wouldn’t mind giving up some forms of reading to machines. “Maybe it’s time to get rid of X-ray [readings], such as chest X-rays. I can’t think of a trainee in the past 20 years who got excited about interpreting X-rays. It never happened.”

Better this less challenging job be done by algorithms, said Dr. Rogalla. “If AI could take over X-ray interpretation, it would be a killer app,” he commented.

For her part, Cindy Siegel said there are several areas where AI could provide incredible value to an organization.

For example, when reading many exams, radiologists are focused on the issue at hand. “But it would be valuable to be able to screen patients and pick up incidental findings that point to near-term or future problems,” she said.

“When you interpret a CT or an MR, you’re looking really at the acute findings. You’re not focused on the chronic diseases, and that’s where population health really comes into play. If you’re able to find something early, then you’re better able to treat it.”

Dr. Siegel commented, as an example, that a radiologist might be reading a lung exam. But if an algorithm, acting as an intelligent assistant, observes that the patient has experienced bone loss and has reduced height, you can treat the patient for this, as well. “In this way, you can reduce mortality and morbidity rates.”

On a related front, Dr. Chang stressed that AI could be invaluable in addressing the inefficiencies of the healthcare system.

“The majority of that inefficiency comes from things that should be automated but haven’t yet been automated.

“In other words, the tedious parts of the workday that require little in the way of sophisticated thinking. These little things fill up our days, not allowing us to think about other things that are critically important.”

Dr. Siegel commented there are many tedious tasks he could think of that are ripe for automation. “Finding rib fractures on a CT of the chest, looking for lung nodules, and in ultrasound, trying to track and count thyroid nodules.”

He said, “There are many things we do that end up becoming really repetitive.”

How far away from solving this problem is AI?

“I’d like to say that most of the technical challenges of building an algorithm have been done,” said Dr. Chang. “If you give us enough data, we can build a tool that does what you want.”

At the same time, he cautioned that this doesn’t account for everything. “Once we have a good tool, do we have good integration that leads to good workflow? That’s what we don’t have, and that’s where work needs to be done,” he said.

Dr. Chang is confident, however, that “it’s just a matter of time before we iron out some of these important details.”

Even if machine learning and other forms of AI can take over some or many of the functions of a radiologist, the panelists agreed that these technologies will never replace radiologists – at least in the short-term or medium-term. “Fifty or 100 years from now, who knows,” opined Dr. Rogalla.

He observed that with the greater automation of radiology, with solutions like PACS and time-saving tools, there has been a reduction in the cost of imaging. And when costs go down, he averred, demand goes up – which has been the case in diagnostic imaging departments.

“In the short-term and medium term, we will need more radiologists,” he said.