The State of Radiology AI in 2019

For years, one continuously swirling question in radiology has been whether artificial intelligence (AI) has become sophisticated enough to be used in clinical practice-and the most dreaded question of all: whether it is advanced enough to unseat the practicing provider. So far, the answer has been “not yet.”

And, for those waiting with bated breath, the answer is still no-and, it won’t be any time soon. But, according to many industry experts, there continues to be a great deal of ongoing work devoted to developing tools that can streamline and expedite the daily activities of the radiologist.

“The hype for artificial intelligence is far from what is actually being used as artificial intelligence,” says Alexander Logsdon, MD, an early interventional radiology resident at Nova Southeastern University.

But, while AI isn’t making clinical treatment decisions options, it is continuing to grow and develop, creating inroads into improving workflow efficiency and productivity. And, at some point, says Keith Dreyer, DO, PhD, vice chairman of radiology computing and information sciences at Massachusetts General Hospital, the industry will stop thinking of these tools specifically as AI.

“As the technology continues to advance, we’ll see things like we’re seeing today, such as early diagnostic solutions, to assist the radiologist. We’ll also see things that will enhance acquisition built into the scanners,” he says. “It will optimize workflow, productivity, and patient flow. After a while people will stop calling it AI and will think of it as simply something that makes their work better.”

Current AI Uses

Despite early predictions that AI solutions would eliminate the need for the radiologist, the uses and applications, to date, have been limited. The FDA has only approved 28 algorithms, but several offer the same solution, such as identifying suspicious pulmonary nodules or reducing CT dose.

In addition, Logsdon says, approved AI solutions also exist for identifying intracranial hemorrhage, spinal fractures, pulmonary embolism, pneumothorax, and rib fractures. Other abdominal algorithms awaiting approval focus on pinpointing free air, hemorrhage, dissection, and aneurysm.

Clinical implementation has been stymied, however, because these solutions aren’t readily generalizable throughout all facilities, says Mitchell Schnall, MD, chair of the radiology department at the Perelman School of Medicine at the University of Pennsylvania. Most AI tools, he says, are tested on small data sets that are largely unique to individual facilities.

Related article: Beyond the AI Hype

“Most AI solutions are not trained to work across data sets,” he says. “Most of the products we’ve tested can’t integrate and function in our system, and even those that do usually aren’t able to reach their claimed performance standards.”

This obstacle is difficult to overcome, he says, because many AI companies either don’t have access to large enough data sets to make their solutions feasible in most settings or they don’t have the bandwidth to test the products with enough clinical partners.

What Does AI’s Future Hold?

According to Dreyer, there’s significant room for growth in AI over the next decade.

“Over the next 10-15 years, I feel very confident that we’ll see algorithms that are in limited use today being deployed inside clinical practice for radiologists,” he says. “And, beyond that, we’ll see AI solving things that are difficult to solve today. “

These tools will be better able to completely segment organ systems on ultrasound, CT, and MRI, he says, as well as quantify lesions on previous and current scans, and predict morbidity and mortality from a series of X-rays or CT images.

Schnall also predicts improvements in reconstruction software that is currently designed to clean up images before the radiologist reviews them. There’s currently little confidence in these tools, he explains.

“I struggle with how much credibility to give these products because they’re really using data to estimate what typically is an artifact and what typically isn’t to eliminate what shouldn’t be there,” he says. “And, the entire focus of radiology is to pick up on the unusual.”

Improvements to reconstruction software are anticipated within the next two years, he says.

Is AI Still Scary?

Even as AI products take on more complicated and complex problems, their role is continuing to develop as one of a reliable support network for the radiologist.

“There’s no reason for radiologists to be nervous about artificial intelligence,” Schnall says. “In fact, I’m delighted.”

Dreyer agrees, adding that radiologists must monitor and manage how AI is deployed in their particular institutions, however. Doing so ensures tools function properly and put the patient first.

“AI is going to help the radiologist, the patient, and the referring clinician,” he says. “It’s going to make us more accurate, make the field safer, and help us provide more value than we’re able to provide today.”

Still, Logsdon says, radiology, as an industry, is unlikely to come out completely unscathed from more widespread AI implementation. Some circles continue to think of the specialty as a commodity, as they want cheaper and faster imaging. That desire could lead some referring clinicians to rely more heavily on AI solutions instead of their radiology partners.

“I don’t know if the radiologist will ever be directly replaced by machine-learning AI,” he says. “But, I think AI machine learning could supplement the clinicians enough to the point where the clinicians replace the radiologist.”

It’s also possible, he says, that the need for the radiologist could drop once AI solutions become more readily available in the cloud or are better integrated into the hospital systems.

Navigating the Near Future

Both Logsdon and Schnall agree adopting and incorporating AI solutions will require active work among radiologists. As with most changes, Logsdon says, successfully integrating these tools will require a culture shift in many offices. Many providers still need to be educated on the benefits of using AI, as well as how to use it correctly.

It’s also important, Schnall adds, for providers to learn how to identify when AI errs.

“Radiologists must understand if a mistake has been made, and they must know how to discount the mistake,” he says. “If an AI tool makes too many mistakes, then you’re spending more time discounting the mistakes than you’re reading imaging. You’re trading a headache for an upset stomach.”

To help providers navigate these issues, Dreyer says, the American College of Radiology’s (ACR) Data Science Institute created the ACR AI-Lab to help radiologists educate themselves on AI and practice using the tools. Not only can radiologists access videos and cases that teach them how to use AI, but they can also design their own to test how the technologies work.

Related article: How AI is Evolving in Diagnostic Imaging

“AI-Lab is a place for radiologists and clinicians who don’t know anything about AI to have a place to start,” he says. “By letting providers build and apply models they’ve built, AI-Lab has allowed people to be onboarded into this space where it’s extremely difficult to start from Ground Zero.”

Ultimately, Logsdon says, with continued AI growth, radiology is set to strengthen itself and demonstrate a greater level of value in healthcare.

“A lot of smart people and a lot of money are working on AI. There are some radiologists and residents who don’t think AI is going to impact them in their career,” he says. “I think they should think about that again. If we embrace it, and we become the keepers of images, I think we’ll really enter a second Golden Age of radiology.”