Wait. Will AI Replace Radiologists After All?


Polarizing debate continues to simmer over where artificial intelligence will take radiology. On one side are those who believe AI will—maybe not soon but eventually—render diagnostic radiologists obsolete. On the other are those who see any and all doomsday scenarios as so much science fiction.  

To be sure, there’s plenty of middle ground between the two extremes. But don’t all arguments, if followed to their logical ends, lead to one of them as a final landing point? 

As of early 2020, it’s unlikely a clear debate winner will emerge before some future development makes all points and counterpoints moot. But proponents of both viewpoint categories have much to say that’s worth hearing. 

Robert Schier, MD, RadNet Northern California

For starters, consider the doomsayers. In commentary published in the July 2018 edition of JACR, Robert Schier, MD, a neuroradiologist with RadNet Northern California, described the divide  this way: 

“There are vastly differing opinions [on AI in radiology], from the apocalyptic claim that AI will make all radiologists extinct to the delusional assertion that computers will always merely assist—and never replace—radiologists. Both extremes are mistaken, but the truth is in the direction of the first. … Unless radiologists do things other than interpret imaging studies, there will be need for far fewer of them.”

While he’s not quite predicting a radiological apocalypse, Schier describes himself as an alarmist. (He emphasizes that his opinions do not reflect those of RadNet.) And he does indeed believe that the end is coming. 

“My guess is that in 10 to 20 years, most imaging studies will be read only by machine,” he tells RBJ. “The results will be transmitted directly to the referring physician without input from a human radiologist.” 

This arrangement will produce faster, better and more accurate diagnostics, Schier believes. And everyone will receive equally excellent diagnostic care. 

“There will be no difference,” he says, “between an exceptional radiologist and a mediocre one.”

What’s good for patients will be traumatic for the practice of medicine—and, as AI expands, perhaps for all of humanity, Schier says. But that’s a different article for a different day. 

Shreyas Vasanawala, MD, PhD, Stanford Medicine 

By contrast, Shreyas Vasanawala, MD, PhD, professor of radiology at Stanford, remains optimistic, seeing AI as assistive, not at all adversarial. He agrees AI will change the way radiologists practice, but

AI allows radiologists to pull information that would otherwise be left on the table, Vasanawala explains. “AI enhances the value of medical imaging,” he says, “which is great for patients as well as the field of radiology.”

Resistance Can Cost 

Technological advances have always raised concerns among the potentially affected. Vasanawala points to the fussing and fretting that accompanied the field’s migration from films on lightboxes to images on screens. Having experienced the efficiencies of PACS, “no one is hoping to go back to film,” he notes. 

Falgun Chokshi, MD, MBA, a neuroradiologist formerly with Emory University who’s now on his own as a speaker and advisor, sees AI as another notable milestone on that timeline. 

“As with all technologies that have been introduced into our practice, from cross-sectional modalities and PACS to voice dictation-based reporting, those radiologists who have adapted and integrated the technology have fared much better than their more obstinate counterparts,” Chokshi says. 

Falgun Chokshi, MD, MBA, independent speaker and advisor

Since AI technology is inherently neither good nor bad, he adds, many negative reactions to it are misdirected.

“Its eventual uses, direct consequences and collateral effects are all based on how humans use it,” he says.

The question is how the human/computer collaboration will play out. And only time can supply an answer.

Known Unknowns

Nobody really knows what AI is capable of, Maciej Mazurowski, PhD, of Duke points out. “I think a significant disruption is a real possibility,” he tells RBJ. “However, many things have to align for this to happen.”

For one thing, AI needs to perform at the level of a radiologist for a broad range of tasks. “I believe there are signals showing that this is possible,” he says. “However, we don’t know yet if that is true due to a still limited clinical validation of algorithms.”

In an opinion piece published in the August 2019 JACR, Mazurowski staked out a spot in the aforementioned middle ground, suggesting AI could well displace more than a few radiologists. 

Maciej Mazurowski, PhD, Duke Health 

“[S]uch disruption is a real possibility,” he writes. “Although the radiology specialty has shown an astonishing ability to adapt to the changing technology, the future is uncertain, and an honest, in-depth discussion is needed to guide development of the field.”

The radiology specialty “was born of technology and has grown around technology,” Mazurowski continues. “It has shown the ability to evolve. It is possible that AI will transform radiology into a substantially altered specialty in which a human specialist will still play an important role.” 

Permanently In 

Pointing to the mass of scientific literature published to date, Chokshi forecasts AI will increasingly focus on classification and rudimentary prediction of abnormal findings in medical images. But that doesn’t mean diagnostic radiologists will completely stop tackling those tasks themselves. 

Human readers “use multiple levels of nuanced perception and interpretation of images that the machines have not been able to rival to date,” Choksi says. “I believe—and some studies have shown this—that human-machine assistive hybrid teams are better for specific findings than just machines.”

Vasanawala’s Stanford colleague, Curtis Langlotz, MD, PhD, has consistently insisted that radiologists’ role is eminently augmentable but will never be replaceable. At conferences he’s often noted that very few airline passengers would ever board a plane with no human pilot even though autopilot technology already handles upwards of 90% of every flight. 

And in commentary published in 2019 by Radiology: Artificial Intelligence, Langlotz makes this compelling point: 

“We often compare AI algorithms to radiology experts based on the ability to identify a single disease or a small set of diseases. But that oversimplifies things. A radiologist may be looking for numerous conditions and … anything else suspicious that might show up in a patient’s test results. That requires a human.”

Eventually Out 

Shier, the alarmist, also sees AI as assistive—in the short term.

He’s certainly correct that AI systems are constantly getting smarter and faster, steadily approaching a point at which their rewards may yet outweigh their risks. 

“It may help to imagine these systems not as a collection of circuits in a console but as an army of fellowship-trained radiologists with photographic memories, IQs of 500 and no need for food or sleep,” Shier writes in his JACR essay. 

AI will initially make radiologists more accurate and efficient, he explains to RBJ. Then AI systems will take over the reading of certain simple cases, expanding the range and complexity of cases they interpret on their own. 

Whether it takes another 20 years or 50, the day will arrive when machines “won’t need us,” Schier says. “The evidence that computers will become better than humans in interpreting all imaging studies and in reporting those interpretations to other physicians is, in my view, convincing.” 

He sees no fundamental physical or theoretical reason why computers won’t be able to do radiologists’ jobs faster, better and cheaper than humans.

“I think the profession of radiology, as described by physicians who interpret diagnostic images, is going to be mostly gone at some point,” Schier adds. “Radiologists, if they exist, will have to do something else.” 

What about the argument that AI will open the field to new screenings and diagnostic tests currently based in non-imaging protocols? 

“Perhaps it will,” Schier allows. “But what role will radiologists play? It may expand the field of diagnostic imaging and at the same time contract the field of diagnostic imagers.”

On the bright side of Schier’s vision, humans and machines together will continue to be better than humans or machines alone—for a few more decades or years, anyway. 

Tireless Assistant 

If Schier is overly alarmist, and humans remain in the picture throughout the lifetime of rads working in 2020, what will these human-machine assistive hybrid teams look like?

Vasanawala surely speaks for many when he answers that AI will allow radiologists to work at the top of their license and to “focus on the most value-added and rewarding parts of the job.”

At present, radiologists don’t have enough time in the day to get through all items on their clinical worklists—or, at least, with minimal distractions drawing eyes from images. AI can help carve out time for clinical concentration by automating dictation and report editing, gathering clinical histories, finding relevant prior imaging, making careful measurements of lesions and correlating current lesions with those from multiple prior studies. 

Radiologists can then increase their level of attention to detail on the complex cases, integrate the full clinical story and look up the relevant literature, Vasanawala says. They can spend more time diagnosing diseases, preparing for and participating in multidisciplinary team conferences and consultations with patients.  

Beyond Reads 

Current discussions of AI largely focus on automated interpretation of images, such as chest x-rays. But there also are opportunities for using AI to optimize practices’ fiscal performance.

Radiology practices, like other businesses, have many moving parts—intersecting workflows, bottlenecks, key performance indicators and significant resource and financial allocation and uses. “AI can help optimize many aspects of the radiology business as long as the users are cognizant of the technol-ogies’ limitations and capacities,” Chokshi notes. 

He adds, however, that, like any technology, “AI is only as good as its intended use.”

Vasanawala, too, stresses “upstream” benefits of AI. He rattles off a handful of these: improved patient scheduling and protocoling of exams, better management of radiation dose and reconstruction of higher quality images, prioritization of studies to read and automated extraction of quantitative information imaging that is currently too time-consuming to perform routinely. 

“AI will enhance the value of radiology in healthcare, but the entire medical imaging department—schedulers, technologists, nurses, administrators—needs to embrace it,” Vasanawala says. 

Some of this is already happening “under the hood” in ways that aren’t always evident. 

And new areas are emerging. One of these is automated segmentation, Vasanawala points out. For example, in cardiac imaging, fully automated, highly accurate segmentation of ventricles is leading to greater radiologist efficiency. And with that is coming greater reproducibility of quantitative results. 

Aligned Interests 

Evaluating AI’s most exciting opportunities and most daunting threats, Chokshi says, comes down to asking two questions: Does the technology help improve patient care? And does it “help maintain our own sanity and mitigate burnout?” 

He’s wary about all the promises being made on AI’s behalf around efficiency and productivity. 

“I hear a lot of talk about using AI to make radiologists faster and more efficient,” Chokshi says. “I would be very cautious about such dubious notions of what AI should be doing for us.” 

A better opportunity, he says, is using AI to become more specific about interpretation, triage acute findings and decrease the variability of reads for different kinds of studies.

“The goal should be alignment of better patient care with better radiologist care,” Chokshi say. “The threat of AI is not the technology itself. It’s using it to squeeze in more work with diminishing long-term rewards.” 

Misaligned Incentives 

Regardless of one’s position—pessimist, optimist or somewhere in between—radiologists must come together to address the relevant issues even more than they have already. That seems to be a widely held opinion, based on the literature and the interviews conducted for this report.

The need for profession-wide cooperation around AI has been coming to the fore as the topic gets attention at regional, national and international conferences. As a result of the presentations, exhibits and discussions, Vasanawala, points out, viewpoints are evolving rapidly.

Schier suggests there’s no time like the present to build on the momentum. He hopes one of the major academic journals serving radiology will devote a full issue to radiology thought leaders reasoning together on the question “Will intelligent computers ever replace radiologists?”  (He fleshes out his own opinion here.)

But it’s not just radiologists who need to come to the table. Rads will evolve and adapt to AI being integrated into workflows and practice settings, Chokshi says, while developers, researchers and regulators work with radiologists to tap AI for streamlining workflows and optimizing outcomes. 

Mazurowski calls on radiologists to collaborate with AI developers in industry on ensuring that priority No. 1 remains improving patient care while holding the line on costs.

And that’s not just a technology challenge, he warns. The best-case future scenario for patients is having AI interpret medical images with the highest possible accuracy and the lowest possible turnaround times—all for the lowest possible price. 

However, Mazurowski says, even if AI proves capable of superhuman performance across a broad spectrum of radiology tasks, that vision may not come to fruition. 

“One of the most challenging hurdles with implementation of imaging AI in clinical practice in the U.S. will be the complex and suboptimal structure of the healthcare system,” he says. “The incentives in the system are often misaligned with the interests of the patient.”

Here to Stay 

Now is the time to figure out how to align systemic incentives with patient interests, the experts agree. 

“As with any new technology, there’s a period where the technology becomes better understood and refined, leading to a progressive uptake in that technology,” Vasanawala says.  

Langlotz contends that the only radiologists who should worry about losing their jobs to AI are those who don’t use AI. (He makes his full case in a piece published online last May in Radiology: Artificial Intelligence titled “Will Artificial Intelligence Replace Radiologists?”)

Meanwhile Schier exhorts radiologists to embrace AI as readily as they can. “It will make your reading of studies faster, more accurate and less stressful,” he says. “You will provide better patient care, be more valuable and do a better job.”

It’s reasonable to expect those benefits to continue building for the next 10 to 15 years, he says. 

“Embrace AI because it will be good for your patients,” Schier concludes. “Embrace AI because it will make you a better, more efficient and happier radiologist. Embrace AI because if you don’t, you will be replaced by radiologists who do. Embrace AI because, even if you refuse to embrace AI, you will nevertheless be embraced by it.”

That series of imperatives may not settle the debate over radiology’s eventual fate at the hands of AI, but the soundness of its wisdom would be hard to argue against. 

Trimed Popup
Trimed Popup