Skip to main content
Episode 3
Duration 24:54

Will AI Transform Radiology?

Howie, a Yale emergency department radiologist, and Harlan discuss whether artificial intelligence will replace human radiologists—or help them do their jobs better.


Harlan Krumholz: Hello, everyone. Welcome to Health & Veritas. I’m Harlan Krumholz.

Howard Forman: And I’m Howie Forman. We are physicians and professors at Yale, and we’re trying to get closer to the truth about health and healthcare. What’s something that’s exciting and attracting your attention this week, Harlan?

Harlan Krumholz: Well, it’s hard not to pay attention to this big change that’s occurring at the NIH. Just this week, Francis Collins announced his retirement from the position of directorship of the NIH. Here’s a guy who’s served for 12 years under three presidents and done with remarkable distinction.

I’ve had the pleasure of working with Francis in many different venues. We were on the board of governors together at PCORI. I was on the advisory committee to the director for many years with him. And I just saw a person who is truly remarkable. Someone really dedicated to science and public service, really a model of integrity and humanity. He stood up for diversity within the NIH. He stood up for women, for young investigators. He tried to write some wrongs. By the way, he worked with the Lacks family to make sure that the NIH did right by them. He was able to translate complex concepts into really understandable nuggets that he was able to convey to the public. He was a terrific advocate on Capitol Hill. Anyway, you can see I’m a big fan.

It’s an exciting moment in science. We’ve gotten a lot of progress quickly in the pandemic, and I think that’s gotten people excited about the prospect for other diseases. NIH, Francis leaves it in good shape, but it’s still, it’s at a moment in time where there’s likely to be a lot of change. And so I think we should all be watching carefully who’s nominated and what direction might they bring them. But we’re unlikely to get someone quite as able to communicate across the aisle. Look, anyone who was nominated by Biden, Trump, and Obama is a special individual. He bridged faith and science. He’s a man of deep faith. I recommend his books. They’re really quite interesting. And we’ll see what happens next, but it’s been quite a ride, I think, having him at the helm of NIH.

And what about you Howie, what’s on your mind?

Howard Forman: We are blessed to have a huge nurse practitioner market, so to speak, in the United States right now. We’re seeing about, I think about 35,000 new nurse practitioners completing their programs each year. There are about 325,000 nurse practitioners in the United States right now. And we are fast approaching a time where there will be more nurse practitioners practicing in primary care fields than there are physicians practicing in primary care fields. And in my opinion, this has been a wholly good thing, but it is not something that we’ve really had a national conversation about. And there is friction that continues to exist between physicians and nurses on this particular topic. And I’d love to have more conversations about this.

Harlan Krumholz: So, Howie, let’s get to the main course here today.

So, there’s this question that’s floating around medicine. You’re a radiologist. Are radiologists going to exist in 10 years? I mean, we’ve got all this AI and machine learning and seems like every day they are advances in pattern recognition and computer vision. Are we going to be needing radiologists in the future?

Howard Forman: That is the big question. And it is something that for the last decades, literally you can go back to around the time, I think, I was born that people started using computer-assisted diagnosis of one type or another, and that’s advanced over the years. And now with AI and machine learning, the ability to help read studies or in some cases actually read imaging studies like CT scans, chest x-rays, ultrasounds is becoming more and more of a reality. And so about five years ago, I was asked to write a small piece on this. And I said that I thought it would happen, but it would happen a lot more slowly than people thought. And I think that’s been the case. I think we’re seeing slow progress made in machine learning, AI, helping do radiology diagnoses. And from my point of view clinically, right now, this has been a wholly good thing.

And if one day radiologists are not necessary, in my opinion, that would be a great thing. It’s not like we need to employ physicians in one specialty or another. Our goal should be to provide the best care we can to patients at the lowest cost. But I don’t see radiologists going away anytime soon. And in fact, there’s a shortage of radiologists right now. And I think some of that is because people have been frightened away by AI, but for me working with it day in and day out when I’m in the emergency room— it’s actually been pretty impressive.

Harlan Krumholz: Well, OK. So let’s unpack a bunch of that stuff because there’s so much in what you just said. First of all, shortage of radiologists—isn’t that self-imposed? I mean, you guys could double the number of residency slots. I mean, that’s just a matter of creating a market where you guys can make a lot of money and there’s a scarcity of talent out there to do the work. So, I mean, if you worried about that, you could solve that problem tomorrow, right? Aren’t there people you turn away who would be qualified to do radiology?

Howard Forman: Yeah. I mean, it is always interesting. I talk about this a lot about, how do we end up with shortages in medicine? And we’ve created so many hurdles for somebody to be able to be qualified in any specialty that we do sort of control it, but we also create our own problems. So if a residency program doesn’t expand, if an individual residency program doesn’t expand, they’re contributing to a shortage, but who is looking over the best interest of the public in terms of how many radiologists we need? Even if Yale were to expand 10 extra slots, for instance, that’s a minuscule impact and yet there’s no single body.

Harlan Krumholz: Would the world will be better with 10 times as many radiologists? I mean, we have lots of qualified people applying to medical school and only one in three can get in. And yet we turn away many really great people. And we talked about this before, about the criteria for medical school admission. I mean, there are people, especially atypical candidates, we turned away. I mean, maybe I’ll just save this for a topic for another day, but this is one thing about the shortage.

But the other thing is you said that you were in the emergency department and I’ve always wondered, why is it that you have to be in the emergency department? I mean, why aren’t we having sort of centralized radiology farms? Just given the cloud, why isn’t this a gig economy where basically anybody in any time zone can read a radiograph? So why do you even have to be in the emergency department?

Howard Forman: Yeah. And we don’t. So we have, and the pandemic accelerated this, but we don’t have to actually be in the emergency room. I’m one of the few people that only reads from within the hospital. There are many people now who read from home, read at their convenience. We have people who have workstations in two locations, so that even if they’re on vacation, they’re available to help out in an emergency. So we do, we have distributed imaging and we currently have sufficiently capable and happy people working night shifts and evening shifts and off hours of one type or another. But we are very capable of doing time shifting. We’re capable of having somebody read from a remote region. There are regulatory reasons that make it a little more difficult to do that, but we…

Harlan Krumholz: Yeah, but those regulations were from a different era. I mean, no one contemplated the current situation. Why doesn’t Yale just simply outsource this? And why aren’t there just large numbers with high levels of quality control and multiple readings of particular images so that we really bring this into the modern day.

Howard Forman: So, I think we are. We at Yale, because we have really pretty large scale. We have around 100 radiologists. At the level of scale that we achieve right now, we’re able to have subspecialty reading across the enterprise, so that our inpatient and outpatient studies will be read by one group, emergency patients are read by a different group, specialized neurologic imaging—whether it happens in the emergency room or the inpatient service or the outpatient service—are read by neuroradiologists. Trauma CTs will be read by trauma radiologists. We’ve continued to specialize, and we have the scale to be able to have sufficient people on at all times to be able to cover that. Many other places really can’t and what you described is absolutely beginning to happen, where various practices are learning that the only way they could provide the level of quality that they want to provide is to outsource to one or another of teleradiology companies.

Harlan Krumholz: Let me ask you now about this AI thing, because as I think about it, one of the great things that AI does is pattern recognition. And one of the great applications of it is in what people call high-dimensional data, which is the kind of data that’s produced by our most advanced machines that produce imaging. And by the way, the data that’s being produced by these machines goes beyond what our eyes can perceive. I mean, there’s a lot of information that isn’t fully captured within two or even three-dimensional images. And so the computers are able to take into account all this information that’s generated and then be able to produce essentially signatures—this is consistent with X, Y, or Z. And not just that there’s a nodule and is it slightly bigger or smaller, but actually be able to, with much greater precision, be able to tell us about trends over time and so forth.

You use this word “assisted.” Instead of replacing, it may be that this is about augmentation of capability. It’s about putting the pilot, the radiologist, still the human being, in a position where they’re being supported by lots of instrumentation and output in the same way that a pilot in a plane is. And that makes the care, the reads, the evaluations much better, higher fidelity, but they’re still as a human being incorporating this information with the help of the computers, but maybe being able to work faster and presumably better. I see this as coming on really quickly. And I don’t know, I think it may be that the radiologist’s role may be evolving from just simply a manual process to one that’s integrated within a platform that is providing information and then ultimately guidance, right? There is this piece about what you’ve seen, what does it mean and how best can it help that individual patient, which is a complicated thing. But I don’t know, how quickly do you think it’s going to turn? And then what are the implications for training the next generation of radiologists?

Howard Forman: Yeah. At the moment, what we’re seeing is that one of the highest-value use cases is for AI to highlight which cases need to be read most urgently. So if a patient has had a large vessel occlusion or a patient has had a cervical spine fracture or a pulmonary embolus, having something bring that to the top of the list more quickly, saving even 15 minutes or 30 minutes in the time to diagnosis, might have a positive impact on the patient, particularly when it’s a reliable system. In addition to that, it’s very useful for quality control. Radiologists are human beings. They make mistakes; they make interpretive mistakes. An extra set of eyes has always been of benefit. It’s been proven for decades now that if a second set of radiologist looks at an image that there’s a higher likelihood of picking up more findings and perhaps having an impact on patient care.

Harlan Krumholz: You did a study of this in the ED, didn’t you? I mean, just remind me again, what did you find when you looked at that? And didn’t you review all those CT scans and see whether there had been something missed?

Howard Forman: Yeah, over 20 years ago, we found in the single digits, but high single digits, were missing significant findings on patients in an emergency room setting—this was early in the evolution of an emergency room clinical environment—but something significant, not necessarily management significant, not necessarily changing patient outcomes, but something that you would’ve liked to have included in the report wasn’t included in the report. And that was something we showed a long time ago. And it’s still true today that when you’re reading studies, more eyes find more findings.

Harlan Krumholz: Yeah. And so that makes sense that the AI can layer on top and be helpful. I guess I’m just thinking about the way that this is going to happen, because it’s going to collide with the business model for radiology. And the question sort of is going to be, how is this going to be integrated into practice? And then also, what is it going to take for us to have confidence in the kind of output that these models make?

To me, the promise of it is so great because you don’t need to find someone with gray hair who’s seen a bunch of images in order to ask their opinion. I mean, it means from almost the first day someone’s out, they can tap into the wisdom and knowledge that everyone else has had because these systems can get smarter over time. Just like the first-time pilot is in a very strong position because of the instrumentation and the information around them that helps them become as good as the best. We want a radiologist on day one in the job not to say, well, this is my first day and I need to get better over time with more experience, but to be assisted. By the way, this happening in surgery and a variety of other venues as well.

But what’s it going to take for us to be confident about the output and how are we going to manage the business model issues that are firmly embedded? I mean, people have expectations of what their job is and how much they’re going to make and the kind of revenue that’s generated. I mean, this is quite disruptive.

Howard Forman: And I think that’s one of the reasons why, as you point out, people have not been anxious to expand the workforce, because the uncertainty of what the future looks like has scared people away from it for a period of time. And so I think the fear for some programs has been, if we expand our slots, we’re just going to get maybe a lower quality medical student applying to our program. And so people have kept it tightly wrapped into the numbers that they have right now. I do think we have a shortage in the short run. And as you know, it takes about seven or eight years between the time somebody gets interested in radiology to the time that they’re capable of practicing radiology. And so that alone introduces a substantial lag that we’re dealing with right now. People have been scared off by AI. And I think that has contributed somewhat to the shortage.

Harlan Krumholz: But some of that’s anti-competitive. I mean, we keep making the training programs longer and longer. Presumably this technology might enable us to get someone ready to be a radiologist faster.

I was just wondering why’d you decide to become a radiologist?

Howard Forman: I was inspired by the sort of fund of knowledge that the radiologists seemed to have, and their ability to very quickly integrate an image into a clinical picture to make a diagnosis. They always seem to be the ones coming up with the unusual diagnoses that nobody else was thinking of and pulling that proverbial rabbit out of a hat, but it was also a place where you get to see the use of technology evolving. From the moment you enter the career, it never stops. It’s always moving very quickly.

Harlan Krumholz: You remember the day that you wanted to become a radiologist?

Howard Forman: Oh yes, I do. I met John Landis who was a pediatric radiologist in the neonatal intensive care unit at the Schneider Children’s Hospital, when I was doing my pediatric rotation, and watching him do a head ultrasound on a newborn baby, and seeing the anatomy laid out on a screen real-time with his hands, moving that probe in ever so slight angles to show the lateral ventricles and the brainstem and so on. I remember just thinking, this can’t be possible— and it just inspired me. And he was also just the brightest and nicest man I ever met in medical school.

Harlan Krumholz: When you see the medical students today, you’re mentioning that they’re considering what the future of radiology is. Are you hearing people express doubts once they consider radiology as a field about whether it’ll exist in the future?

Howard Forman: I think a few years ago, I was. There was a lot of fear it was just going to go away. I think what we’ve seen with AI is that it really is assistive and supportive of being a radiologist. It’s not replacing us anytime quickly. And it does continue a trend that has occurred since the beginning of my career. And that is our productivity goes up substantially every year. And it’s because of the various technologies that surround us that allow us to do that. It’s the only field I know about, I’m sure there are others, but where the payment that we’ve gotten over the last 25 years for each thing we do has just gone down over time. But our salaries haven’t gone down that’s because we’re just doing more per unit of time.

Harlan Krumholz: Do you think radiologists deserve to be some of the most highly compensated people in the entire medical profession?

Howard Forman: You and I have had this conversation before. I don’t—I think it’s a great field and I spend a lot of time with my undergrads talking about why it ends up being a highly compensated field. But I certainly think that we do not pay our various specialties according to their clinical importance and the effort that goes into it. It’s basically paid because of the artifacts of the system that we’ve built.

Harlan Krumholz: Yeah. We may need some redesign there for everyone, not just for radiology.

Well, I think that’s really interesting. Let me just, finally, I mean, this isn’t meant to be a question to answer for you, but since you’re in radiology, do you see any downsides by the way? Have you noticed any issues with regard to this that you think could lead to unintended adverse consequences?

Howard Forman: Yeah. I mean, the risk is that you become so reliant on it that you’re not putting in the effort you should. Most of the algorithms we see right now augment the accuracy of a radiologist, but it presumes the radiologist is looking as well. If a radiologist were to become lazy and just say, I don’t need to look for this because the AI algorithm is so good, it’s as good as I was anyway, I don’t have to spend time on it, you could have a quality issue there. Even as the algorithms get really good, they’re still going to be better by having the extra set of human eyes as well.

Harlan Krumholz: Well, that’s great.

Howard Forman: Harlan, what’s something that inspires you or maybe keeps you up at night right now?

Harlan Krumholz: Maybe the same thing that’s been on a lot of people’s mind lately, which is, let me just say, Facebook, and not just because of Facebook itself, the company, the decisions, the controversies, the whistleblower, but the whole thing.

The idea that we really are going to need to grapple with the potential benefits and harms of large-scale social networks, power that’s concentrated in large companies. And then this outage that occurred that went across Facebook, Instagram, WhatsApp. Here, this company that is a technology company that has the geniuses of the world running it and yet, what the heck happened there? They said it was caused by a faulty configuration change. What does that even mean? Can we really get to the root cause of it? Because if Facebook can’t keep themselves up, if they’re going to have outages, and by the way, we see outages from time to time with almost every major tech company. So this isn’t merely about Facebook, but as our reliance on the cloud and as technology grows, I think that we have to be ever more thoughtful about what vulnerabilities that dependence creates.

So it’s the whole package. It’s the ethical part, it’s the regulatory part, it’s the technology part. It’s gotten me thinking a lot about our reliance and use of these platforms and the ways that they can be optimized for good. And then how are we thinking about the redundancies in the system that are going to protect us so that we can be sure that this just doesn’t simply happen. Our whole monetary systems now are based on electrons moving from place to place. And so society has really in a very rapid period of time come to be in a very different position around these technologies.

And anyway, that’s one thing that had my attention this week. What about you, Howie? Is there something that’s inspiring you or keeping you up at night lately?

Howard Forman: I’ve been a little sad or disheartened about just watching Congress do the Build Back Better legislation. And it’s a reminder—I spent a year working in the Senate, so I’m well aware of the fact that there are two things that you don’t want to watch being made:one is a sausage and the other is legislation. And this legislative process can be quite daunting and really break one soul at times when you see the strength of lobbyists lobbying very hard to dismantle things that they think are against their self-interest. And specifically watching the pharmaceutical industry lobbying very, very hard about drug pricing legislation for Medicare and watching dentists, among others, lobbying against expansion of Medicare for dental care for our elderly and disabled. But the reality is this is what goes on all the time. It’s just right now, we’re seeing it up close and personal and hoping that we actually do get something that bridges the divide and can advance the interests of the people who need help the most.

Harlan Krumholz: Do you have a solution for that?

Howard Forman: I don’t think I have a solution for that. I think I’m just in a position of acknowledging that this is the imperfectness of our republic, and this is the way we function right now, but certainly others have proposed solutions, including regulating the moneyed interests in our elections and regulating moneyed interests in Congress. And these have generally failed to some degree on First Amendment grounds.

Harlan Krumholz: Yeah. I’ll just say the thing about it that bothers me is I see an inexorable journey towards more fragmentation and more confrontation. If an external threat like a pandemic can’t bring us together, I’m not sure what will. And most great societies fail by internal problems rather than external problems, or many do. And I worry about that for us. So we need to continue to be a constructive force for good. But yeah, it’s disheartening.

You’ve been listening to Health & Veritas with Harlan Krumholz and Howie Forman.

Howard Forman: So how did we do? To give us your feedback or to keep the conversation going, you can find us on Twitter.

Harlan Krumholz: I’m at @hmkyale.

Howard Forman: And I’m @thehowie. That’s T-H-E-H-O-W-I-E.

Harlan Krumholz: Health & Veritas is produced with the Yale School of Management. Talk to you soon, Howie.

Howard Forman: Thanks very much, Harlan. Talk to you soon.