Artificial Intelligence in Radiology: The Game-Changer on Everyone’s Mind

AI’s Impact Will Be Monumental—Will Radiologists Go Along for the Ride or Be Left in the Dust?

One day soon, machines powered by artificial intelligence (AI) will interpret even the most complex clinical images as accurately as today’s most experienced radiologists. These robot radiologists will automatically generate final reports, uniformly structured and with no need for preliminary reads. Their interpretations will take into account not only all relevant prior imaging exams but also patients’ complete medical histories. The virtual doctors’ doctors will work 24/7/365 and have no falloff in diligence due to fatigue, monotony, interruptions or distractions. What’s more, patients, referring physicians and federal regulators will have no less confidence in the competence of nonhuman medical diagnosticians than air travelers have today in computerized autopilots—which, by the way, already fly your plane more than 90 percent of the time you’re in the air.

At least, that’s the popular narrative on artificial intelligence AI in radiology. Believed and circulated by a widening circle of observers, enthusiasts and startups seeking investors, this sci-fi-comes-true story seems to be everywhere of late.

Many radiologists say such a scenario will never come to pass. AI will augment rather than replace their profession, they maintain, and that goes for as long as medical imaging itself is extant.

A third perspective stakes out a middle ground. Scratch the “soon” from the above doomsday prediction, say forecasters of this third way. One day, they say, AI will indeed replace radiologists—but that day is so far in the future that it’s not worth seriously preparing for by anyone alive in 2017.

Regardless of where the overblown hype leaves off and the sensible forecasts pick up, here’s what matters right now: AI and its offshoots, machine learning and deep learning, are already changing radiology. Research and development projects are hopping. Radiologists are talking to vendors who are listening. And the activity is oriented toward the future, granted, but it’s also shaking up the present. AI-attentive experts within radiology all agree on that much.

More Time for Proving Value

In a recent paper published by the Journal of the American College of Radiology (JACR), digital radiology pioneers Michael Recht, MD, of NYU Langone Health in New York and R. Nick Bryan, MD, PhD, of the University of Texas in Austin, envision a not-so-distant future in which AI becomes a routine part of radiologists’ daily lives, making their work more efficient, accurate, satisfying and valued (J Am Coll Radiol. 2017 Aug 19).

The authors’ educated guess of what’s around the corner includes the possibility that, just 10 years from now, no medical imaging study will be reviewed by a radiologist until it has been pre-analyzed by a machine. This pre-analysis will help separate truly urgent items on image-interpretation worklists from those that can wait, for example, while also performing routine reading tasks such as quantification, segmentation and pure pattern recognition.

Throw in automated data mining of patient histories stored in the EHR, and AI will, Recht and Bryan expect, free up radiologists to “perform more value-added tasks, such as integrating patients’ clinical and imaging information, having more professional interactions, becoming more visible to patients and playing a vital role in integrated clinical teams to improve patient care.” Being the doctor’s doctor more of the time, and adding value.

Interventional radiologist Seth Berkowitz, MD, director of informatics innovation at Beth Israel Deaconess Medical Center in Boston, seconds that. He’s currently working with a startup whose algorithm will comb through hundreds of head CT slices in search of abnormalities at a pace no human could ever come close to replicating. If it finds anything, it will flag the exam as a high priority on a radiologist’s worklist.

The startup’s founders “are not trying to replace radiologists,” Berkowitz says. “They are trying to improve outcomes by doing something the radiologist can’t do.” The underlying idea is to allow radiologists more leeway to channel their energy toward “proving that what we do is a lot more than just translating pixels into words in radiology reports.”

[[{"fid":"23469","view_mode":"media_original","type":"media","attributes":{"height":773,"width":537,"alt":" - AI-morethanjust-CAD20","class":"media-element file-media-original"}}]]

Point: Radiologists Will Not Be Replaced

Berkowitz co-authored a paper published in JACR summarizing the 2016 summer meeting of the American College of Radiology’s Intersociety Conference, which annually brings together 53 radiology societies (J Am Coll Radiol. 2017 Jun;14(6):811-817). The organizing theme of the meeting, he says, was clinical data science and its application to the daily practice of radiology. “One quote that kept coming up was: If you as an individual radiologist could be replaced by a machine, you probably should be replaced by a machine,” Berkowitz recalls. “Yes, looking at images and describing findings is a very important part of what diagnostic radiologists do. But, at a much broader and deeper level, diagnostic radiologists define care-management plans based on information.”

The JACR paper underscored his point. “Radiologists will not be replaced by machines,” the authors wrote. “Radiologists of the future will be essential data scientists of medicine.”

To this, Berkowitz adds: “In some ways, that paper was a call to action to radiologists. We were saying, in effect, ‘Get out there and tell industry the kinds of things you want to see.’ Otherwise, you’ll have industry supplying what they think you want. And the other worry is that you’ll have hospital administrators who come out and say, ‘What I would really like would be an algorithm that maybe is not as good as an expert radiologist but as good as a bargain-basement radiologist—at half the cost.’

“Obviously,” Berkowitz says, “that would be a huge disservice to our patients.”

Counterpoint: It’s Only a Matter of Time

One AI watcher who believes the technology will eventually replace radiologists—and may not take a lifetime to get there—was a radiologist himself before leaving clinical practice to work as a C-suite hospital executive. Jason Kelly, MD, MBA, chief medical officer of Sky Ridge Medical Center in metropolitan Denver, first made his prediction public in early 2016 in Forbes.

Does he really foresee the day when that Forbes piece—entertainingly but ominously headlined “The Robot Is In, And It Will See You Now”—is looked back upon as prophetic?

“I think it will get to the point where AI is reading so many studies that a human being can’t actually check them,” Kelly says. “It’s going to be largely automated. The way an automobile manufacturer does quality assurance on the machines that are assembling the cars, it’s going to be like that. A human being will read one out of every 20 CT scans that the AI has read first. Humans will provide the QA process, but eventually AI will be doing the vast majority of the reading without any human intervention.”

Kelly points to the growing demand for medical imaging, which is constantly multiplying the number of studies to be interpreted, as one driver of the inevitable rise of the fulltime radiological robots. Emergency room patients getting scanned before being seen is another sign of things to come, he adds, as are skyrocketing scan and slice counts. Top all that off with the relentless push to squeeze out every dime’s worth of unneeded expense, and the picture starts to resemble an earlier era in technological history.

“You could say that the car didn’t completely replace the horse and buggy, but it kind of did,” Kelly says. “And now the self-driving car is right around the corner. I’m not saying this is going to happen in the next few years. I would think that anyone who is currently a radiologist will probably encounter this in clinical daily practice at the tail end of their career or it may happen after they’ve retired, but if you’re a high school student thinking about going into radiology, you might want to think about going into computer programming instead.”

Kelly further predicts that, 20 years from now, medicine as a whole will have metamorphosed, as the synergistic drivers of technology and economics simply won’t be slowed, much less stopped. “It’s going to be much less labor-intensive and much more reliant on data,” he says. “Medicine is going to consist of collecting data on the patient, figuring out what the right algorithm is and going down that route.”

All the Action Is in Augmentation

The next 20 years will see momentous change in medicine, but numerous obstacles loom for any AI developers who are of a mind to speed to market radiology algorithms with any whiff of potential to displace radiologists.

There’s the FDA, for instance, which is far from being ready to consider, much less approve on a premarket basis, AI solutions aiming for autopilot-level diagnostics. Even winning 510(k) clearance to expand legacy CAD products for new AI-based indications will be onerous, as AI is more sophisticated than traditional CAD by orders of magnitude. There are the patients, nearly all of whom are going to need a lot of familiarizing with nonmedical AI before they even think about giving thumbs-up to the notion of having “robot readers” as doctors. And then, of course, there are lawyers who are always on the hunt for perpetrators of medical harm. Profit-driven software developers would make juicy targets on whom to pin legal culpability, and they know it. (Recht and Bryan flesh out all these hurdles in their JACR study.)

And then there’s the AI development that is actually in the works. Mark Michalski, MD, executive director of the Center for Clinical Data Science at Massachusetts General Hospital and Brigham & Women’s Hospital, both in Boston, gets to see, touch or at least hear about most of it.

[[{"fid":"23470","view_mode":"media_original","type":"media","attributes":{"height":458,"width":537,"alt":" - the-college-is-allinwithA","class":"media-element file-media-original"}}]]

“We have the good fortune of working with great researchers, and nearly all the machine learning algorithms that are being assessed, developed or even conceptualized are for radiologist augmentation as opposed to replacement,” Michalski says, offering as examples algorithms that not only prioritize worklists and flag anatomical abnormalities but also calculate probability metrics for critical findings.

“As powerful as these tools are, they still must be checked by a human to get to the levels of confidence that we need,” he says. “That may not always be the case, but it is the case today. Radiologists will be guiding the development of these tools for the foreseeable future, and these tools are not for final diagnosis.”

Michalski adds that radiology is far from the only specialty hashing out how best to use AI. “Practically everyone is talking about this,” he says. “We are having similar discussions with pathology, radiation oncology, cardiology and neurology. In fact, every ‘-ology’ is having this same discussion. It just so happens that radiology is once again at the forefront of a technology revolution.”