The new radiology operating model | Radiology Rewired | Season 2 Ep. 1

What does it actually take to implement AI in radiology — at scale, in the real world?
Not in a lab. Not in theory. In a 4,000-radiologist practice covering hospitals across the country.

In the Season 2 premiere of Radiology Rewired, host Dr. Vivek Singh (Neuroradiologist, MUSC) sits down with Dr. Ryan Harvey — practicing radiologist, Clinical Assistant Professor at Florida State University, and President of Radiology Associates of Florida — to unpack the real mechanics, human friction, and future of AI-powered radiology.

From ambient dictation and multimodal vision-language models to care coordination and AI governance, this conversation doesn't shy away from the hard questions: Are we training residents wrong? Will AI displace radiologists? And what does a true AI platform actually look like vs. a pile of disconnected tools?

This is the conversation every radiologist, radiology trainee, and healthcare AI leader needs to hear.

🎙️ SPEAKERS

Guest: Dr. Ryan Harvey, MD
Radiologist & Clinical Assistant Professor, Florida State University
President, Radiology Associates of Florida (Rad Partners)

Host: Dr. Vivek Singh, MD
Neuroradiologist, MUSC
Host, Radiology Rewired Podcast


🎙️ Listen on Apple Podcasts and Spotify
🔗 Learn more: https://www.rapidai.com/podcasts 

Transcript

SPEAKER_01 0:09

Welcome back to Radiology Rewired and welcome to season two. This season, we're taking a deeper dive into the many dimensions of radiology, exploring different facets of the field through a range of perspectives. But we're also widening the lens beyond radiology itself, bringing in voices from across the healthcare ecosystem, from C-suite to physicists, surgeons, even legal experts, to unpack how AI is rewiring healthcare from every angle. In today's episode, we're doing a deep dive on what it takes to actually implement AI as a radiologist and how better integration can start to connect radiologists more directly with the clinicians who depend on them. And then we'll zoom out. What this means as imaging volumes keep climbing, how teams keep up without burning out, and why the way we implement AI matters just as much as the technology itself. And there's no better conversation to have this with than Dr. Ryan Harvey. Dr. Harvey is a practicing radiologist and president at Radiology Associates of Florida, one of the largest practices with radiology partners. He leads clinical operations across a high-volume multi-site group while still reading cases on the front lines. He's lived the full evolution of AI and radiology, from early NLP tools to today's more integrated platform-based services, and has been directly involved in evaluating what actually alleviates the pressures facing radiology. He brings a grounded, real-world perspective on how AI is impacting radiologists day-to-day, how it's changing expectations and productivity, and how it's influencing everything from care coordination to patient perception. Dr. Ryan, thanks so much for joining me.

SPEAKER_00 1:52

This is a real privilege. So much has changed so fast in the last couple years. I think we just all radiologists should sit at a table like this and just take it in for a minute, like pause and talk about what's happening, because uh working today is a lot different from uh working even just a few years ago. So lots to talk about.

SPEAKER_01 2:10

Yeah, absolutely. I mean, we were talking a little bit about stuff before, and and you know, you kind of entered the workforce in like 2013, 2014 time period. So you told me a great story, but just tell me a little bit about how you ended up here, what your career's been like. You've seen a lot of changes over the past 10 years.

SPEAKER_00 2:25

Yeah, I maybe, maybe my sort of like elder millennial generation had the perfect uh place to stand to see technology transform radiology. When I interviewed for my job, it was uh with one of my still current partners, Dr. Liechtenstein. He was sitting in a reading room hanging a film on a light box. Uh next to him, there was a little microcassette dictaphone recorder that he would hand to somebody who would take it away and transcribe it and bring it back to him. So uh that was the ground floor for me uh at in in Radiology Associates of Florida. And then uh very soon as we scaled and got bigger, we obviously fully embraced PACs, voice recognition. Uh, then, you know, the next shoe to drop was integrated PACs across many hospitals. And uh most lately, uh, and like I said, it's been very quickly changing, AI. AI that touches everything. It's getting deeper and deeper integrated into the workflow. So um I think I was telling you, it almost gives me a whiplash to really wind back the clock 10 years. It's it's changed so much.

SPEAKER_01 3:26

Yeah. So talk me through that a little bit. You know, I work in more of like the kind of larger academic center setting. Um, so I think from my experience, our tools are very slowly adopted. They have to get kind of approved across the system. Um, I think that, you know, corporatized like medical groups like rad partners or bigger conglomerate groups have the ability to kind of implement this stuff a lot quicker and evaluate a lot faster. So take me through what kind of tools you guys are using.

SPEAKER_00 3:51

Yeah, sure. Uh and and it's interesting because I've seen it both ways. We we weren't always a rad partners group. So though there was a time when technology procurement and then trying to integrate it with the hospitals uh was kind of a big lift. So uh obviously the practice is huge. We're a scaled practice, 4,000 rads or something, you'd be count across the nation. Um, that that is not bona fide by itself. Like being big doesn't mean being good. One thing that being big has unlocked uh is innovation and implementation. So um when uh there's a new tool that uh gets procured by our practice, uh, we have a lot of people coming on site to help us use it, uh to integrate it with our hospitals. Uh, we don't have to solve those problems. Hopefully the person that's helping us with that has done it a thousand times. And and honestly, even before that step, it's gone through some silent validation. So I'll tell you about the tools we use, but what's nice about a lot of them is before we spend all this energy trying to use them, because honestly, as a radiologist, there's a cost to getting started with a new tool. It changed anything that touches your workflow and changes the way you've done things for years, um, it's a speed bump right at first before it pays dividends. So uh before we get to that point, uh, there's very diligent uh sort of pre-deployment monitoring. We'll run an A2 an AI tool in the background silently. And let's say it's a tool that is supposed to aid in detection or interpretation. Uh, we'll compare its output to the ground truth, the radiologist reports. And only when it gets above that threshold where it's good enough that I don't have to triple and quadruple check it and it's going to speed me up, uh, that's when it gets deployed. So um that's where we were probably around, you know, 2022, 2023. Fast forward to now, um, you know, we can go into detail and whatever you want, but uh we have language tools that touch report generation. Uh we have measurement extraction tools that help us import measurements from pencil and paper worksheets without having to painstakingly, you know, uh act like we're we were trained to do data entry and radiology residency. Um and then more recently the multimodal tools, and and this is where it you know really feels like you're living in the future, vision language models that uh you know first are sophisticated enough to uh find individual diagnoses and now are sophisticated enough to uh, and this is an investigational tool, it's it's pre-clearance, uh, draft reports based on what it sees uh in entire images. So uh very exciting to see these models progress. Uh scale has been an advantage in turning them on. And uh, you know, not just on the technology side, but you know, imagine you and your partners uh having these really powerful tools changing how you work. There's some help we need in actually implementing them, both among the radiologists. There's a lot of change management involved in anything that touches our workflow, uh, but also with our partner hospitals. Uh a radiology tool that lives on one computer is nice, but it's nothing like a radiology tool that integrates with, you know, hospital A, hospital B, uh outpatient imaging center C. And, you know, that's where having a larger footprint can kind of help in those implementation lifts.

SPEAKER_01 7:16

Yeah, makes sense. So are you guys all for the hospitals that you're currently serving, are these tools like pretty much throughout all the hospitals now?

SPEAKER_00 7:24

Yeah, yeah, they they are. That's you know, the the tool, the environment that we deploy into is a cloud native environment. So what's nice about that is we don't have to ship a CD to every workstation and go through the Windows install Wizard. It's uh, you know, it gets turned on server side. And once the big integration is done with that hospital, once that hospital is on our radiology platform, um, adding a tool is just a matter of getting appropriate security clearances, making sure it's going to play safely with all the sensitive data it touches, uh, and then it turns on. Uh, but like I said, the the shipping and the the IT integration, sometimes that's the easy part, the actual practice management angle, and uh teaching radiologists how to use it, because uh AI is not set it and forget it. These uh these these tools are sophisticated and they require a sophisticated driver, a driver that knows what they can do, but what the limits of what they can do is. And uh it's not obvious out of the box. Anybody who's used Chat GPT knows that it's a revolutionary tool that can sometimes get it wrong. So uh that that's that's an important part of it on the front end, too.

SPEAKER_01 8:32

Yeah, absolutely. So you have all these things kind of getting gradually adopted into practice. What's been the change on the practice side for you guys? Have you guys noticed increased efficiencies and like how are people liking it?

SPEAKER_00 8:42

Yeah, yeah, it's it's interesting. Um the I think the first, our earliest tool was pre-LLM. So pre-ChatGPT style AI, pre-transformer model. It was a natural language processing tool that would look at our reports in real time and surface for us appropriate recommendations that uh an up-to-date radiologist would make based on the findings. For example, if we see an incidental thyroid nodule, that ubiquitous finding on a CAT scan, not all those need an ultrasound. It's it's it's waste for the patient, it's stress for the patient. Some of them probably deserve an ultrasound. There are evidence-based guidelines that that tell us which bucket that thyroid nodule falls into. So this tool would understand that we saw one of these nodules and it would surface us the right recommendation to make. And with a button, we could put it into our reports. So the first change we noticed was not really radiologist satisfaction. It was not even radiologist efficiency. It was what was previously an unmeasurable quality metric. Are we practicing evidence-based medicine? So it was, it was really a value add for the patient and our hospital partners. Now, more recently, we've gotten to, for a radiologist, the good stuff, the stuff that touches our work. Uh some of the generative language tools that we use to generate reports. For example, you know, we might speak casually to the computer in an ambient voice mode about what we see on the image, the way I would be describing to you when I saw on a brain MRI. We were chatting about it. And there's no template there. I'm not navigating fields, I'm not trying to figure out how to fit my thoughts into discrete boxes. Yeah. Just putting the thoughts out there naturally. And then the AI does the work of synthesizing those into a template after the report is done. It's an inversion of the usual model where it's template first, then the ideas, and then a report. It's ideas first, don't worry about the template so much, then a report. That's a kind of tool that uh on the margin improved efficiency, but the real gain was satisfaction and felt stress at the workstation. So especially when we started at first with just the impressions, uh we measured maybe a marginal efficiency increase. Uh but the real win was radiologists just felt like you know, two hours of work was was not as stressful as it was yesterday. So uh that that's a huge win. Uh and to me, that's I don't know how you want to define productivity. I think the the classic definition is uh how many patients is a radiologist taking care of in an hour? Yeah. Output per time. Maybe another measure of productivity is output for input, and maybe there are other inputs besides time, like blood, sweat and tears, stress, and uh lost sleep and uh missed dinners and staying late at the workstation. So uh you know, if a radiologist is getting even the same output, but with less blood, sweat and tears, then that is uh uh an efficiency win in another way. So that was that was the first win was uh less friction, less friction. Now though, as we've really gotten into multimodal models, um this is where the the win is more uh truly affecting efficiency. What what can a radiologist do in a given amount of time? So uh we're starting to to be able to measure that.

SPEAKER_01 12:02

Um so you mentioned like templating and like this, you know, ambient dictation. So just for our lay people listening in the audience, a lot of the times with uh with radiology these days, we have these templated reports. So if you open up a chest x-ray, it'll have lungs and then it'll have pleura and it'll have bones, and next to it it'll have a little section that basically has like certain common findings. So something like uh no consolidation or no pleural fusion, no pneumothorax. So radiologist right now and these days, like you'll get this template that pulls up and you kind of sift through the template, and then you you know add things in as as necessary. So what he's describing is kind of not having a template, and it's just basically you're just describing your findings and it's creating that report for you. So it's really exciting for us because I mean, I know I run into that issue all the time where I'm like, this process falls into like kind of two buckets, and I kind of want to just delete the whole template and just like free dictate. Um, like but again, it just takes more time. So hearing that you can make these just like observations and how that actually affects like the workflow for things is really exciting because you know burnout is like pretty high in radiology right now.

SPEAKER_00 13:05

It is. We we need all our neurons pointed at, you know, being a doctor, making diagnoses, helping our partner physicians and helping these patients. But the truth is, a radiologist right now is not just a physician. We are a word processor. We are sometimes uh a data entry specialist. Uh, we are very often a uh pixel hunter, looking looking for the bright, bright dot in the lung field, for example, to find that small lung nodule. And uh, you know, we spend mental clock cycles on that. And uh that the value they add, if there's technology that can safely and effectively do it for us, you know, I question if that's the value we need to add. Like I would love it if uh I could design a machine that could help me with the data entry. And that's exactly what what AI is doing for us now. So uh I want to spend more of my time in the intellectual sphere of uh in context with the patient's history and their presentation. What is this telling me? And and what useful can I tell the physician that that sent this patient here? And and if I could spend 99% of my time on that, then I think the phrase that gets uh tossed about is we're practicing at the top of our license. Yeah. We're, you know, that that's that's medicine in the in the imaging lens, really.

SPEAKER_01 14:18

And and a lot of these tools have I think helped foster some communication between radiologists and clinicians. I think we've kind of fallen out of favor with some clinicians because a lot of us are remote or you know, you're kind of just grinding through studies. And, you know, I I think that the value of the radiologist back in the day, as more of a clinical investigator, providing and providing like insight in person was really valuable back in the day. And I think we've moved away from that. But I'm starting to see the shift, you know, with some of the tools that, you know, there's live consult systems where you can go through imaging with the clinician live and things like that. Have you guys had any tools like that? Or or have these like improvements in efficiency led to better communication for Yeah, we need to lean into that.

SPEAKER_00 14:59

And and yes, we've uh we've we've dabbled in that area for sure. Care coordination, connecting the radiologists to the physicians that are depending on us. I think there's a latent anxiety among radiologists that as the technology becomes more powerful, the human radiologist might regress from the care loop. And uh I think that's one of the, you know, let's call it controversial conversations or crises even in the specialty. What does it mean to be uh an AI-empowered radiologist? Well, uh, to the extent that AI can bring radiologists more context, meaning the patient history, their labs, their prior imaging reports, more context than we've ever had at our fingertips before and can connect us more closely than ever before and more quickly than ever before to the physicians that order the exams, we're not regressing from the loop. We are becoming more central in it. So uh I think embracing AI technologies, especially integrated technologies, models that are state-of-the-art and frontier quality are great, but if they're not integrated and they don't have agency onto autonomy, uh we're we're leaving value on the table. If we have AI that uh lets us deliver care directly to those patients more, more quickly and communicate with with our uh referring physicians more quickly, uh then we really have added value and we're more central in care. I mean, stroke and pulmonary embolism are two domains already where uh you know time really matters. And we already have some live tools uh at the workstation that uh alert me as a radiologist that there are some stat findings on this exam and have likely already sent some pages to uh the neurovascular team that's gonna cure that stroke. So um care coordination and integration is probably the biggest lift, in my opinion. Uh, you know, uh the these models are getting better and better, but uh maybe maybe it's a tougher problem, or maybe it's lower-hanging fruit, to integrate the the tools we have to uh be visible and active and have agency across what are historically disconnected systems, packs uh our our viewing system, wrists, our information system, the electronic medical record where the patient's medical chart life lives, uh, and then all the systems, not to mention that our our consultants uh might be using. So uh one example is uh you know, moving away from stroke and pulmonary embolism. One communication tool that we've tried was a mobile app that uh connected the radiologist to the emergency department. So there would be ER physicians who'd have a mobile app on their phone, and that mobile app let them uh see when an exam was uh being worked on by the radiologist, when the result was ready. And if the if the physician, the ER physician, was particularly worried about a patient, there'd be a button they could push that would flag the case and bump it up on our work list. So those are the kind of ways to shorten the loop and the distance that's been increasing between the radiologist and the consulting physician. Uh and I think we need to lean into that as a specialty. Docs are not coming to the radiology department anymore to do X-ray rounds. That that was a great, you know, didactic chance to really uh get heads together and do problem solving. But with with PACs and remote technology and the internet, it's allowed us to widen the distance between the other physicians in in healthcare and and what radiologists do. Uh interestingly, it's gonna be technology, I think, that closes the loop a little tighter. So uh it's exciting and very formative, but I think that's the next generation is integrated AI tools.

SPEAKER_01 18:42

Yeah. And just, you know, speaking on that, like I think there's, like you brought up, there's been uncomfortable conversations between that, like basically how our role is going to change in medicine. Um, I know that I get a lot of questions on social media about from medical students about, you know, is this field going to be changed completely? Like, is it worth applying to and things like that? What are you saying to medical students these days? I mean, you guys are right on the front, you're testing out the newest model. So I want to hear kind of your take on all this stuff and how you see things progressing.

SPEAKER_00 19:11

Yeah, and let's state it plainly. The question is uh, is radiology a job worth pursuing when when AI is starting to touch uh just about everything we do? Uh are are we ripe for replacement? You know, I I think the in a hundred years, I honestly don't know. I'm not a futurist what it's gonna look like. I don't I don't know what being a doctor is gonna look like in 50 years. Right now, if you if you are a uh a medical student trying to decide on a specialty, uh I not only think that AI is an advantage for you, this is a great time to uh grab the wheel and use it because you're gonna be empowered as a radiologist using AI compared to, let's say, the last gen radiologist. Uh the new radiology will not be isolated in the reading room, uh making observations based on a clinical context that is often just two words at best at times. Uh the empowered radiologist with AI tools uh will hopefully be wielding a system that is essentially a central brain with a rich knowledge layer of that patient's entire clinical context, uh, with the friction sanded down. So they're using their medical expertise to take care of that patient. Uh and what does that mean? I mean, I think ultimately what the question is is I mean, this sounds great and it's interesting, but is this going to be a viable living for me as a physician? I've got to do something that society needs. Does society need me? Uh, yeah, more than ever. Uh a radiologist that brings more than just an accounting of the pixels on the screen, um, an information specialist, a diagnostic specialist, uh, that is a high value physician. So um I myself am excited about it. I I try to build excitement in, you know, medical students, residents who who visit us for interviews. Uh, and I think it's uh is still, and for my foreseeable future on the horizon that I can see, uh, a very worthy field that AI is is helping. And and you know what? Honestly, we we have to face something. Uh society expects it of our of our specialty. Um, you know, healthcare is too scarce, uh, not accessible enough, uh, and there's always better outcomes we could uh we we could achieve. Um if AI unlocks that, then uh we've got to do it. Yeah I mean uh it we need to add more value uh uh in you know in healthcare. And I think AI is the unlock that we're waiting for to do that.

SPEAKER_01 21:33

The future of radiology isn't more tools, it's the right tools connected. You may know Rapid AI from stroke, but the platform has evolved. Today, Rapid AI is a full body enterprise solution built to support imaging teams across the enterprise. And at the center of it is Navigator Pro, a radiology workspace unlike anything else on the market. Navigator Pro prioritizes your work lists, services AI outputs right where you need them, and generates Generates reports automatically, all within the systems and workflows your team already uses. No new logins, no new context switching, just a smarter way to work. But here's what makes it especially compelling. Navigator Pro connects radiology to the broader care team, putting radiologists in a more central role in patient care decisions, not just reading images, but driving outcomes. Find out how Rapid AI can restore your mental energy and raise the profile of radiology at rapidai.com. There's a huge shortage of radiologists right now. We were talking about before the interview, but our lists are constantly backed up and have these crazy fluctuations. And it seems like there's never been, you know, more of a need for radiologists now. And yet people are like, I don't want to do that. I don't want to, you know, and it's like your job market's getting better. You're it's the best it's ever been. Your job opportunities, like the way you can set up your practice and the way you can work have been better than ever because you are in such high demand. And I think people just see that that demand is just catapulting in like the next in the near in the near future with all these like, in my opinion, sometimes sensational headlines um and just to garner attention. And we all know about this AI bubble just to try to get the funds going, you know. Um and I I feel like it's a very nuanced discussion because, you know, yeah, you have this thing that's going to help speed up the process and help us make make better decisions and be more efficient. But I mean, what about all the new developments on the side of just interpretation, like speeding up protocols and speeding up screening protocols and and improved access and portable imaging units and all these kind of things that are going to be and that are being developed? Like if we talk about imaging explosion now, what's gonna happen in 10 years? Like, how many images are we doing then? Everyone's gonna get a screening brain MRI for whatever, or screening some some kind of study, and there's just gonna be this huge influx of imaging. Because I mean, phase it, we've moved on from the physical exam and we're just taking pictures now. Yeah.

SPEAKER_00 23:57

Yeah, that's right. I mean, I'm not a macroec uh economist, and uh, you know, I would I would hesitate to to really put my nickel down in a forecast, but here, here, since you brought it up, here's my forecast. All right, I love it. Wind the clock forward. Uh radiology volume increases more than more than it has. Uh even the patients already uh getting taken care of by radiologists, we do have a backlog. We can do better on timeliness, we can do better on outcomes, uh, we can definitely do better on cost. I think all of those will improve first. So, yes, I think it'll cost less, hopefully, for a patient to get a CT scan if we find efficiencies in new AI technology. So I think that uh access and cost will go down for patients, but the balance of work that is on the radiologist plate will go up. Now, that might sound stressful to a radiologist, but I think that the volume of work that a radiologist does doesn't mean that the amount of stress or the amount of hours that a radiologist has to uh commit to do that work has to go up. The the promise of of AI is as we need to, we'll be able to take care of more patients safely, more efficiently, uh, with less friction and you know drudgery, honestly, uh perhaps at a you know more more accessible uh cost and timeliness for the patient. But because our volume will outpace that, the the radiologist will will still be gainfully employed and rewarded for uh you know the value that this technology will add. So uh that is my own opinion and my own forecast, but uh I am I'm bullish on on radiology. And it if it makes anybody feel better who's worried about how AI will quote unquote disrupt radiology, there is a lot of slack to be picked up. If if AI does make us more efficient, does it mean right away we need fewer radiologists? I don't think so. There are a lot of unread exams uh that need to that need to get read and uh an efficiency gain would be more than welcome and is not going to put pressure on the workforce uh right away. Uh I mean it kind of gives you a gut check when every once in a while someone counts up uh how many scans are done and how many uh trainees are coming out of training. Uh and then it's even more of a gut check when someone counts up how many radiologists are likely to exit uh the healthcare system in the next decade. The the demographic curve is not in favor of a boundless supply of radiology experts interpreting images. So um, you know, I I think that this particular time is bullish for radiologists, and uh I think that AI is uh a useful boon to what is otherwise becoming a scary shortage. Uh and you know, I honestly I don't want to have a work list that's getting longer. I uh I I won't I'm leaning in to get that down.

SPEAKER_01 26:45

Yeah, and I think that your point about you know, your people in practice um really feeling less mental fatigue when having the same day and reading similar volumes is of course really encouraging because I mean, like you're saying, the way I see it, it's like you could almost see this as we're heading toward a more almost like more golden age of radiology again. Like the golden age has passed, I guess. Those like in the 80s, and you know, we talk about that stuff whether reading 10 10 studies a day and uh you know, just living the life, um, all perseverating over one X-ray. Um, those are the days.

SPEAKER_00 27:17

But uh good at X-rays back in the golden age.

SPEAKER_01 27:20

And in fluoroscopy, yeah, those guys are wizards in there. I, you know, I'm not nearly as good as fluoroscopy as the people who taught me. Um, but I think a lot of it, you know, nowadays the workflow is completely different. You know, you're just churning through exams, trying to read as much as possible. Like you're saying, you know, lists are backed up. And so it it just it just compounds on itself. So, like just uh, you know, for all you guys like who don't have this experience of sitting there in the reading room, love my job and everything, but there are times when the work really does compound in the sense that your ER list is is is blowing up because you had a bunch of traumas and a few different ERs that you're covering, because it's not just one these days, it's like eight ERs and a couple urgent cares and two outpatient imaging centers that were just opened up in the past year, because again, imaging volume and imaging demand continues to skyrocket. So you have this list that should be your main priority, finishing those stat lists. But then because you've been spending your whole day reading ER exams, people are getting scanned 24-7 now. So outpatient scans are getting done, really complex follow-ups that are just sitting on your list. And so they start to build up when you're just focused on kind of trying to clean up the emergency stuff. So, yeah, you know, I think if we're able to use this to make our reporting more efficient, reduce turnaround times, and get the list more compressed, I mean, I think everybody's gonna be happier in the long run.

SPEAKER_00 28:36

Yeah, and and what specialty is more ripe for sanding down some friction than radiology? Radiology, you know, compare us to internists rounding on the floor, and uh, I really value the work they do. I think that's uh it's it's hard work, it's very intellectually demanding. Uh the radiology loop is every 30 seconds or every two minutes, depending on what you're reading, every 10 minutes. You close one exam and the next one is served up for you. We are continually producing. And unless we engineer in for ourselves breaks, it is just by its nature mentally fatiguing. And the 100th CT on the list needs as sharp an eye as the first CT on the list. Every image on every thousand-image exam could have a life-changing diagnosis on it. So, you know, we owe it to our patients to be uh maximally attentive, minimally fatigued. Uh, and that has challenges with a workflow that is really piecework that repeats every couple minutes. Right. Uh so if we have a tool that shaves 10 seconds off of our loop of interpreting and diagnosing an image, uh, that might sound mundane on paper, but it's multiplied times doing that for however many times it takes in an eight-hour shift, for example. Huge dividend. Yeah. Um even if we don't see a measurable efficiency increase in the traditional productivity metric, say RVUs in a shift or exams in a shift, which we do, but even if you took that off the table, the gains of lowering friction, increasing alertness, increasing satisfaction, decreasing burnout, uh that might be harder to measure, but uh the from where I said just as important. Yes. So yeah, I mean, we're the perfect use case as a field for uh microefficiencies that can multiply some of the AI technology is more than microefficiencies. It's it's you know uh enhancing us in in much larger and more profound ways. But uh yeah, we we are ripe for AI enhancement as a specialty, in my opinion.

SPEAKER_01 30:39

Yeah, and I think that, you know, when especially when talking to people in clinical medicine, like I think AI in clinical medicine and like the greater field of of healthcare is I wanna I don't want to say new, but it's becoming way more rampant in the past year or so, right? Like now you have all these EMR companies doing their own algorithms internally, and and those are getting you know immediately integrated and adopted on a large scale because they already exist. So I think we had that experience, except like eight years earlier, right? Like all the steps or all the issues that they're running into, or all the realizations that they're having about, you know, oh, this doesn't do what we want it to do, or this is not useful to the clinician at all, or it doesn't work into their workflow. I think they're having these same conversations and discussions as we do. We're just like five years ahead. So I, you know, and I I tell medical students all this time, like radiology has, you know, had the AI come up for 10, 15 years now. Um, I think it's gonna continue to progress, but these other fields are also experiencing the same thing. So it's not that you know clinical medicine is an algorithm, but there are algorithmic processes in clinical medicine that can also be made way more efficient by AI. And who's to say you're not gonna have a change in the market there? You know?

SPEAKER_00 31:52

Yeah. Uh I I mean, I think it's reasonable to it's not an extrapolation, it's just an observation. The AI uh disruption, if you want to call it that, is not unique to radiology. Uh it's not even unique to medicine. I mean, I think it's gonna touch any and has already touched any knowledge work and you know, increasingly and controversially, uh art work in in society. So, you know, AI is gonna transform society in a lot of ways, but what you described uh to bring it back down to the hospital, uh AI tools being integrated into everything that's on a computer, EMRs, packs, uh to me, that's a call for what I think we're starting to observe. The next wave of AI is integrating AI so it's not 20 different widgets. Yes. Um and I think right now what it is is there's there's a separate chat model built into every app from the EMR to the food ordering app on my phone, for example, who wants to have uh an AI uh agent telling me uh, you know, what I might like to order next. Uh what we don't want is the radiology interface and the physician interface to be like the space shuttle where you have 20 different lights blinking at you and 50 different levers to pull. It'd be nice if it was just one steering wheel, one system that you could get your hands around. It would serve you the right information when you need it. Uh it could flag you uh the uh the acute findings when it finds them. It could uh extract the data entry tasks out of your hands and uh put them put them right in your report. It could take the template out of your brain and just do it on the back end so you can just talk to it. Um and it all works as one system. Uh integration also solves the downstream problem of how do you get the AIs to talk to each other? Uh how do you get your language model to interact with your vision model? How do you get your vision model to not just tell you what it sees in text, but to show you what it sees overlaid over an image? So that that's the brass ring is AI that's not only capable on the model level, but that is agentic and integrated at the systems level. And the way you described it right now is kind of the the lived experience where we are, a bunch of layers on top of each other that don't necessarily communicate. Um and it's the big investments in platforms. And you know, my my practice is is party to that. Mosaic is a radiology AI native platform that uh is is aimed directly at answering that noise and turning it into something that sounds more like music. Uh and I know there are other entrants uh that that have platforms just like that. That's what's gonna move the needle. That's what's gonna make it feel less stressful because the worst implementation of AI for a radiologist is an implementation that just feels like the treadmill is going faster and you're running faster. Yeah, you're doing more, but you're not really going anywhere. Uh so um fortunately, I I think that we we are about to crest that wave of model integration, not just model capability.

SPEAKER_01 34:52

Aaron Powell What have been the biggest like pushbacks that you've received or your group has dealt with uh with all this AI adoption?

SPEAKER_00 34:58

Yeah, interestingly, they haven't been uh the the biggest ones aren't technical. Uh you know, by the time a model touches a radiologist, it's been through the the the technical filters, it's been through these safety filters. If it's uh ready for clinical use, it's been through regulatory review. So uh they're not technical. They are often uh, you know, they're in the leadership or management domain of change management. So uh let me give you an example. Radiologists uh are experts and are very proud of our output. Our output, the physical form of it, is a report. It's a it's it's a few paragraphs, an organized report in language that we select. Uh generative AI can unburden the radiologist from producing a report that has a specific form or specific words or specific language, and it can let the radiologist just give their ideas, and then an AI tool can actually put words, uh precise words to them. Well, uh that's not the best for every radiologist. Some radiologists are are quite possessive of the language, the precise language that they want to invoke in a report. So uh there's a bit of a psychological shift that radiologists are undergoing as they're moving from uh, you know, traditional last-gen practice to AI unlocked enhanced next gen practice. It is not just, you know, moving from a four-cylinder to a six-cylinder car. It's changing the mental loop of reading a case. And uh any change has friction. There's no frictionless change. Uh so the biggest issues with deploying at scale have been around uh how do you set up everybody for success? I mean, we were talking about scale earlier. That's that's one more domain where scale helps. Uh we didn't just install the package and say, have fun, Rads. Yeah. We had uh the developers themselves come on site, sit on Teams meetings with us, uh, sit behind the workstation, watch what we do, show us how to succeed, show us new best practices that we would otherwise have to just find through trial and error and creating this new style of reports. And uh that that there's no law of physics that says that happens automatically. So um, those are the kind of human-embodied frictions that that we see with the new technology. Um there's institutional and regulatory friction that can be a challenge. Uh AI technology has its own challenges from a security perspective and a data security perspective, cybersecurity. Uh you know, we have to make sure that uh protected health information doesn't leak somewhere where it's gonna put somebody's lab results in a random Chat GPT query one day. So there has to be guardrails around that. Uh and then as we've already discussed, the integration question is friction. Uh it's one thing to integrate among uh a holistic radiology platform, but to integrate those tools with external platforms like EMRs, that requires cooperation and that's another friction piece. So yeah, it's uh won't sugarcoat it. The these great technological tools, there's still a lot of work to be done to actually uh after the switch is flipped to make them effective uh you know as they're designed.

SPEAKER_01 38:10

My assumption is that it's typically the older, more seasoned radiologists that have a little bit more reservation.

SPEAKER_00 38:18

Yeah, you know, interesting assumption. Uh there is a cohort of radiologists who uh struggle with chains because they've their their wires are just soldered in. They've done it this way forever. But there are just as many of seasoned, experienced rads that love the unburdening that some of this technology does. On the other hand, uh, I'll give you an example. We have a few younger rads who are extremely tech savvy. They have already tuned some of our legacy tools to work really efficiently. They've, you know, incorporated scripting tools to make our voice recognition software do exactly what they want. And they've created this whole individual, uh, you know, hastily put together system that works for them. Well, this system disrupts that. The new AI systems, uh generative AI, it doesn't require so much building and uh uh deliberate construction and deliberate consideration. It's supposed to just work based on natural input. So that's a change for them too. So in some ways, uh it's it's bimodal. You know, we have the experienced rads who are gonna have to learn new tricks, and we have uh some very tech savvy rads who they're moving to another tech savvy solution that clashes with what they're using now. So uh, you know, it you you can't, I don't think you can just predict uh what what the response is gonna be in a big cohort just based on demographics. Uh I think you can predict that the the management and the deployment, the change management will have to be very deliberate. And practices that do it right are gonna invest very highly and very deliberately on the front end. They're gonna front load the communication, the training, and most importantly the feedback loop. Uh I'll share one more example of uh where I think you know being in a big practice has helped me personally, at least my experience with these tools. The traditional way that software improves that we use as radiologists is uh we make feature requests that hopefully will eventually filter their way up to uh the developers of this software, which are third parties. Uh we hope that our requests will make it into uh their changelog for a future release. And then we will wait and hope that a year from now, when the next 2.0 update comes out for a piece of software, that the new feature we want will appear. Uh with internally developed tools, the feedback loop is there's really no time or distance there at all. The the dev team and the radiologists who are on the front line, they're in the same meetings. Uh they have each other's cell phone numbers and email addresses, and the iteration is very rapid and unconstrained by the usual uh extramural bureaucracy that you can run into. So um change management is is just as important as the technological expertise in getting these things to ultimately move the needle for rads and patients, I think.

SPEAKER_01 41:12

So you talked about how you know you have increases in efficiency, um, kind of improved burnout in the group. How has that, you know, in a very competitive recruiting environment, has that like incentivized people to join the group? Like, oh, you know, because we ask now in job interviews, or you know, I haven't done one in a while, but everyone I know is asking, like, oh, what AI tools are you using? And everyone wants to know, just like you used to ask what kind of packs you're using, because it you use it every day, it's gonna change your job, it's gonna change your your burnout rate and stuff like that. So have you noticed any difference in the recruiting? And uh is this something that's coming up a lot with potential candidates?

SPEAKER_00 41:47

Yeah, for sure. Uh of course everybody wants to know about your tech stack. And uh this, what is very uh frontier technology, emergent technology, uh piques a lot of interest. It it is a differentiator, I think, if if you're a practice that it has leaned in and has uh a set of capable tools that work well together and are well integrated, I think that's a great sales point. But I'm gonna give you a very honest answer. Um I think right now at this stage, it's icing on the cake. And really the uh the differentiators are still the cake itself. Yeah. And the cake itself for recruits, I think, is uh, where am I gonna live? Uh is my work meaningful? Uh how enjoyable is my work? How stressful is it? And given those inputs, is the compensation fair for that equation? Does it balance for me? Now, the AI on top of that can be a tiebreaker, but but here, right now in 2026, I think the needle is moving to where AI will kind of become the cake. So uh the platform-based integrated AI will change how satisfying it is to be a radiologist and work at the workstation. It will change how stressful it is to be at work. It will change uh the kind of compensation uh that a radiologist can accrue. Uh and that's where I think it won't just be icing, it'll kind of be table stakes where uh, you know, new radiologists who will increasingly be exposed to, I hope, AI in training in a judicious way that still builds a strong foundation uh in radiology for them. I I hope they'll have awareness of the AI tools that are out there uh and an expectation that they'll be close to the state of the art in their ultimate practice. So um I think we're not quite there where ri where where the AI deployment is a deal breaker yet. Yeah. I think we're pretty close to it though.

SPEAKER_01 43:35

Yeah, makes sense. And you touched on trainees, um, and we talked a little bit about medical student kind of reservations these days, even though you know we have this kind of stagnant pipeline of radiologists every year. With all these new AI tools, when you're getting auto populated reports, flagged findings, like do you are you concerned about the future radiologists like baseline understanding? of all this stuff because if you have all these assistive tools, how good is your interpretation if you don't have them going to be? And if something goes wrong, what happens then? You know, I think that's always like my question. It's like the fallback there. Who's your fallback guy if if your outputs are wrong, if no one knows what they're doing?

SPEAKER_00 44:15

Yeah. If uh if the AI server decides to die uh uh Saturday night at uh 2 a.m and your resident is covering the ER, what happens then? So there definitely is a wrong way to integrate AI into education. Uh it can be a crutch uh and you could train a fleet of radiologists who are AI operators without the fundamental knowledge. I don't think that's going to be meaningfully how uh residency programs incorporate AI. Uh it's something that my own practice who staffs University of South Florida uh are, you know, on the precipice of uh really carefully considering now because our our AI deployment is is now uh you know the frontier of our deployment is now going to start touching uh our our training environments. Um AI absolutely needs to be part of education. Think of it like a new modality uh to to artificially excise it from the uh radiology process in the name of training a good fundamental uh medical skill in radiologists will serve the underlying medical skill of radiology, but it will be a disservice for that rad who will have to use AI tools uh when when they when they graduate so it has to be integrated uh like a new modality AI literacy uh needs to be not an elective it needs to be mandatory so that the new radiologist will know what these tools can do but also know what they can't exactly do. But it cannot be a shortcut for uh building interpretative skills, building detection skills, uh and there will have to be mechanisms uh to preserve those, whether that's gating certain tools at certain levels of the residency so that you don't have Gen AI vision tools, for example, right, as a as a first year RAD one. You know, and and and this deserves much longer than a pithy answer. It has to be very carefully considered and it it could be done wrong but I have faith that the academic institutions are are going to take it very seriously and and also give it the the diligence that it requires and the attention it requires. We you know we can't shut our eyes to it. It's a great question though.

SPEAKER_01 46:24

Yeah I mean I I I think about it as just kind of setting up uh you know the way we we used to learn in radiology residency is like you you dictate the exam, the attending is in the background and then they basically go over the exam with you or at least they change your report and then you're able to look at that to see what the ground truth is, what they thought. So then the concern becomes like if you already have your answers at the back of the page, you know, like what's the resident really doing they can just scroll two more slices and be like oh it's negative like so we really need to be able to have them work just on the raw values alone and then use AI to almost verify their outputs or something like that is what I was thinking to maintain that sense of training.

SPEAKER_00 47:02

Exactly uh that those are exactly the kind of things we'll have to consider. But I I also think we could acknowledge that what they're going to learn in practice is going to change. If I had to guess for example I think that developing very rigid search patterns to detect the pixels that are out of place in an X-ray, I think that maybe that will occupy a smaller mind share in education for AI-powered radiologists. But they'll also have to learn something that you and I did not have to learn and that's uh if you have more clinical context than just a one-liner patient history, can you give a better differential diagnosis and recommendation? If you have at your fingertips lab results, consult notes, which you will in the in in platform-based AI, which you do now in platform-based AI, uh there's probably a call for you to use a higher order of you know intellectual processing and diagnosis to render a more meaningful diagnosis. So maybe search gets less important and integration and analysis becomes more important. It's a skill we use, but we were constrained by the the silos that the information lives in. If AI can only do one thing, uh I hope it's information integration.

SPEAKER_01 48:14

That that's a great low-hanging fruit that uh residents will benefit from the we're going to be doing our own clinical correlation finally.

SPEAKER_00 48:23

Yeah no yeah no no cop-outs anymore.

SPEAKER_01 48:25

Yeah correlate clinically uh yeah all your internal work in the mirror and start like look you have all the information yeah Dell that's that's that sounds really exciting and I think it like we were talking about earlier then it then it does like it's just this weird balance right you have like this side almost deflating radiologists and you have these articles and this new CEO talking about stuff and then you have other sides of it where you're like oh well we have this new tool that lets me communicate with clinicians more often and now I can you know say what I really think over like the video and be like I think it's way less likely this. And then you think about yeah like you're saying bringing in really pertinent information from the EMR incorporating that to your impression. So maybe the resident like for all you guys listening who go into radiology residency you're less concerned about like the Fleischner criteria like the Fleishner guidelines and the glossary and how to you know what qualifies as tree and bod nodular that's all going to be done for you and yes you have to recognize that stuff but you're not sitting there dictating those findings yourself. But what you will have to be better in is taking context and using your interior knowledge and then taking those findings and really integrating it into the clinical picture like you're saying I think then your impressions could change dramatically right because you're like based on this this this this not just the radiology imaging like based on the labs and the fact that they had cancer 12 years ago and you know they got seen by an outside hospital like but all that stuff can be pulled in now and it's going to be ready for your review. Because right now our workflow is if I want to find out more about the patient, I have to pull up the chart, click through notes, sift through notes if I want to know about this biopsy if they click pathology, find the date all these different things take time and they you know significantly reduce the efficiency and and lead to burnout too if you want to make that really really nuanced judgment in your impression. So having that at your fingertips I think is really exciting.

SPEAKER_00 50:16

It's very exciting and I think the radiologists who kind of see that where the field is going have every reason to be optimistic and I think that's what it's going to look like. And I think a radiologist that pictures radiology within the narrow boundaries of how it's practiced today or honestly yesterday where uh we don't know too much about the patient except what we see in black and white here on the screen uh we're gonna tell you exactly what we see and deliver our report to you and uh now now it's a you problem. Yeah that that kind of model of radiology I think there is some anxiety about what what is the value proposition for that kind of radiology when there's so much information integration and AI enhancement.

SPEAKER_01 50:50

So yeah I see it exactly like you stated 3D reconstruction has become essential to modern imaging but the process is still slow, manual and fragmented techs spend valuable time recreating reconstructions, outsourcing information to labs and they introduce delays and inconsistencies and radiologists are left waiting for the information they need. But there's a much better way to do things. Lumina 3D from Rapid AI fully automates high quality reconstruction of the head and neck, eliminating the manual burden on your team and the unpredictability of outside labs. Consistent outputs are proven to reduce reconstruction time by more than 24 minutes and enhance radiologist diagnostic accuracy. Lumina 3D gives imaging teams better results delivered to their workstations without the weight or the variability learn more at rapidai.com do you see the like demand for screening populations and stuff like your screening exams do you see that you know catapulting up into the air like in the next few years based on just you know shorter imaging times maybe like uh dose reduction protocols like you're able to get lung cancer screening with like the dose of an x-ray now and and you have better communication and outreach programs and you're going to get your reminder from an AI prompt that your CT chest scan is is coming up like how do you see that stuff transitioning in the future?

SPEAKER_00 52:15

Yeah I mean the we can we can feel those forces already uh some of them are ahead of the evidence like whole whole body MRI screening for example uh you know uh screening is upon us uh whether we like it or not but what what's gonna make it cost effective and evidence-based yes shorter imaging times lower radiation doses machine learning based MRI acceleration protocols that's all gonna make a screening MRI a little more feasible as a conversation to have from a cost benefit perspective and uh exactly even with established protocols like lung cancer screening I think the arrow of progress is more inclusive screening. I think that the guideline changes have favored screening more and earlier uh even in the non-imaging modalities like colonoscopy screening uh earlier. Uh so I think you're right on the money that that's a trend we're gonna see and it's gonna be enhanced by uh all these AI accelerations on the imaging acquisition side, but also on the patient notification side. Uh you know what the biggest barrier to uh lung cancer screening right now is not that there aren't enough patients that meet the criteria. It's that for some reason with the guidelines we have, we're still missing most of the population to get in to inbound them into screen into a screening program. So uh yeah it's coming our way. And by the way, uh do you want to be in a radiology uh world where screening volumes are quadrupled and we don't have AI wind at our back to to do the patients the give the patients the quality they deserve with this increased volume. So uh this this is all going to be a necessity for us.

SPEAKER_01 53:45

Yeah I just see it as all like kind of shifting toward just so much more reliance on imaging for everything. And obviously you have your AI that's going to help in in interpreting inter interpretation, but I think the volumes are just going to continue to skyrocket for us. And I mean I I think it's up to this technology to help us bridge that gap and it's really up to us to adopt it and use it properly.

SPEAKER_00 54:05

Yeah you know I when when we have our meetings looking at our volumes I always say you know there are there are more patients out there that uh need to be taken care of than we can take care of. We would love to partner with every hospital in the state we could, but uh we can't uh you know there there there's a real, I think honorable mission behind unlocking new efficiencies in radiology. Um access to care is not where it should be. And uh you know if we if we can improve that with with these technological unlocks then then good on us. I think I think it's part of our mission honestly.

SPEAKER_01 54:37

You mentioned screening MRI I got to get your take on some of these private companies that are doing it at there was I'm I'm sure you heard there was a recent uh lawsuit affiliated with one of these where a pretty high grade stenosis or narrowing in the car audit was I guess somewhat visible on MRI everyone knows that it's not like tailored to look at that in particular. But now this company is getting you know sued and everything like that because it was a patient had a huge stroke just a few months later. What is your kind of take on all that and how do you feel about the space?

SPEAKER_00 55:08

Yeah I mean I I try to stick to the evidence. We do whole body MRI protocols for for some established use cases, not generally screening for healthy population per se, but right right now screening for um you know patients with certain syndromes that might make them more prone to have cancers but I think there's a real and valid conversation about general whole body screening. Yes, we can see a lot more. We have the technology and pretty soon it'll be cheaper to do it as it progresses. In radiology though we are so acutely aware of the externalities and the cost of seeing stuff that ultimately doesn't matter. Interminable workups for tiny pancreatic cysts every two years for the next 20 years, you know, uh biopsies that have a non-zero risk of morbidity and harm to a patient for uh you know a thyroid nodule that you know might harbor some misbehaving cells, but but might just stay put for the rest of the patient's life. So uh that's a question that has gotten ahead of the science because it's available to consumers. Yes. Uh and uh maybe that is a call to more rigorously you know examine screening use use cases and revisiting the cost-benefit calculations as the product we provide is higher value and more accessible for patients. So you know that that is not a a service line that my particular local practice has leaned into but around my larger national practice I know that that varies a little bit from from site to site. What about you? Have you guys done any uh walk-in screening MRF?

SPEAKER_01 56:50

No, we have not we have not touched anything close to that. Um a lot of our more screening proliferative processes have been mobile MRIs or mobile mammography units going to like more underserved areas in the state because again it's an academic institution. I think their main mission goal is to A expand um and B maintain somewhat of a margin and then C try to get the outreach going. And so that's kind of where we've been targeting things on our end.

SPEAKER_00 57:16

And that's a safe space. I mean that is that is more than supported in the literature and that yes it's screening but uh I almost think we need a a different word for the healthy population screening that doesn't really have a a long track of evidence. You know there are anecdotal cases for for whole body MRI screening oh yeah my dad found a renal cancer that you know who knows what would have happened 10 years later. True true uh individuals make decisions based on you know their individual outlooks individuals don't see themselves as an a small n in a large n population so uh you know the I the the consumers in the market is really driving this more than you know the ivory towers and and and careful research. So uh we'll see what happens when the research catches up.

SPEAKER_01 58:00

Yeah and I think like you brought up you know the the whole incidental finding type thing leading to a very long and prolonged workup. I think some people when they read about these new technologies and stuff they don't think about that side of things where like you're saying we see this stuff all the time we call them incidental lomas meaning like just a little thing there. It's probably nothing accidentally found it. Yeah we accidentally found it and like we've all done like you know dictated those scans from the ER where they're looking for something in particular it's you know an older patient hasn't been seen by a doctor in 10 years and then we find everything you know every little thing in your list of follow-up items is 10 items long. So I think there's a you know a great side where you're like you're talking about you find that renal cell carcinoma but then the other side is what about all this other stuff you're seeing and then I think the biggest thing with this whole lawsuit that's going on is like I think it's showing the need for transparency between the the the corporation and the consumer and being very explicit about what this can and can't do, you know, and if they're able to do that and you know and and show that this this works in certain modalities that's great. But I this whole catch-all kind of scan stuff, I think that's where the misinterpretation occurs for the public.

SPEAKER_00 59:14

Yeah. You know I I haven't done research in this area but I I tend to believe there are probably some more corners of useful screening that will we'll end up having some validation. I I think we need to have the conversation and and have a lot of diligent research there. I also think that the more sophisticated our technology gets, the more we can scare people with what we find. It is true um you know when when people look in the mirror in the morning and they're getting just a judge of like how do I look today? How's my health? It's going to be a lot different if we have on-demand, full body like see inside yourself. Do you want to see what your pancreas looks like today? Do you want to how would you like a a look through your colon? I mean uh it's a brave new world. Yeah yeah for sure.

SPEAKER_01 59:58

So in my experience and you've talked about this word platform and integration um I wanted to delve a little bit deeper for our audience in terms of like what that exactly means. We talked a little bit about how in the past we've been doing single point solutions for radiology where you have this plat this this app that's looking at this specific thing and we'll run it in the background. Then we'll get this app to look at this specific things and you end up having a lot of like segmented outputs um different things to check and like you mess mentioned like different little messenger systems within the app so it just feels very not connected um and I think that has also led to reduced adoption of some of those solutions. So can you tell me about like what exactly a platform is and what you guys are doing over there that's a little bit different than what we've been doing so far traditionally.

SPEAKER_00 1:00:48

Yeah so uh to describe what how I define a platform, let me describe the opposite case uh of like you like you said it four or five different let's say AI tools from different vendors they each flag a finding or give an output and they just kind of dump it in the radiologist's face or screen. And then it's up to the radiologist to integrate all that and to figure out what to do with it. A platform uh is ideally single sign-on. So ideally when I sign on to my workstation all of my tools I've signed into uh the tools aren't operating in a silo. So a vision empowered tool uh directly gives input to a language powered tool and changes what appears in my report rather than a vision powered tool just showing me a flag and then waiting for me to make the appropriate changes in my report. Let's go one dimension more sophisticated. Language and vision tools that talk to each other but can orchestrate my workflow it sees a finding it drafts it in my report uh and before I've even opened it it elevates the case on my work list because it's an acute finding that needs my attention right away. So the orchestration level of agency not possible if all these tools are separate siloed products with separate siloed outputs. One tool that does multiple things, one platform of of many integrated tools that do multiple things uh but work in unison. Uh let's expand it even further beyond radiology uh a platform can cross the specialty boundaries to where it's its value doesn't just accrue within the walls of the reading room uh it alerts the uh clinical physicians of the finding immediately through their uh tools which are part of this platform that institutions might invest in so uh a platform really harmonizes what is otherwise a bunch of nifty tools that just require someone else to put the puzzle pieces together. So it's unified, it's integrated and honestly it's worth considering as a whole separate axis of uh AI sophistication along with model capability. The intuitive axis where we look at uh AI capability is how good is the model? How accurate is it in finding the head bleed? Uh can it play with words or can it also interpret images? Is it multimodal? But the second axis to review a tool or a suite of tools or a platform is how integrated they are in the system. Are they just giving flags? Are they feeding into your report and your viewer tool? Are they communicating, providing some care coordination with pulmonologists and neurosurgeons and the ER physicians uh and is there enough agency in the in the AI uh platform to put all those pieces together taking what is should otherwise be a a rote integrative process out of the hands of the radiologist so the radiologist can stick to practicing medicine and not you know coordinating all the pieces. Radiologists should be integrators, not coordinators. The coordination should happen within the platform. So that's a long answer to say it's not a bunch of bolted on uh add-ons to a a piece of legacy software it's software purpose built to do all this stuff uh from the ground up.

SPEAKER_01 1:04:26

No, that's a that's a great explanation I you know you can tell why that really moves the needle for radiologists because you know you don't want to be like the mental task that it takes to take this output, this output, populate it, integrate all that stuff. It it adds a lot of friction like you're saying and having that stuff all integrated and kind of synthesized for you I think is a huge reduction in friction for your outputs. So we've talked um a little bit about all these different programs that people are using and the you know new vendors and stuff like that. How do you guys go about ensuring that you can trust the outputs from any given vendor um and you know what are the relations relationships like with the vendors?

SPEAKER_00 1:05:06

Yeah uh trust is is is the key. Um if a generative tool is going to make a radiologist more efficient and decrease their headache, not increase it, they have to trust it. If they have to triple check everything uh every piece of output that a tool is giving you it's basically giving the radiologist two exams to interpret in one the actual image and then the AI output. And trust is a challenge with uh you know transformer-based generative AI because those are not algorithms those are uh very sophisticated neural networks that that produce output that might be deterministic but it is not predictable. So if if an AI, for example, thinks that a patient has heart failure based on a chest x ray, it might not be able to point an arrow at exactly why it thinks there's uh heart failure on a chest X ray. Uh it might be very Accurate at making that call, but if you want to ask it why do you think that, explainability of AI is not necessarily uh automatic. So there has to be a lot of effort into transparency in AI so that AI can explain why it thinks what it does. Uh it's also true that an AI tool that has gone through FDA review and hopefully has some studies and some pre-market data on its performance characteristics. That's not enough for the lifetime of that AI tool. Uh AI tools over time, let's fast forward a year, they're going to drift because they were trained on a certain set of images. And a year from now, maybe imaging technology has improved a little bit and images start to look a little different. Is that AI model going to perform precisely the same on, you know, a data set that's not perfectly generalizable from its training set? It's a fact right now that these large models are trained once with billions of dollars of expense, and then they're kind of frozen. They, you know, you can change what they do by um giving them clever engineering system prompts, but you cannot change the model itself without retraining it from scratch, at least the way I understand current large language and large vision models. So uh post-clearance surveillance uh is supremely important. There must be, in radiology at large, and hopefully in every radiology practice that trusts AI models, some uh post-implementation surveillance, periodically check its outputs against ground truth, uh, have ideally committees and radiologists that are looking at this, and be willing to throw a kill switch if something drifts too far. I mean, that's patients are going to expect us to do that. So uh fortunately, uh, you know, many uh many vendors do this. They do post-market surveillance. They're usually required by the FDA uh to do uh some post-market evaluation. Uh and and in fact, radiologists have been instrumental in uh advocacy efforts to make sure that the FDA keeps those guardrails where they need to be. Uh radiologists need to be involved. If there's going to be tools that are autonomous, boy, there better be really great post-market data continuously that that proves the trust. So uh governance is is a great question, and it's why using these tools is not like downloading Microsoft Word and then using it for 10 years. Uh you know, as the population changes or drift, so may the the model performance. So um trust is not automatic, it has to be earned, and uh it takes diligent work to do it.

SPEAKER_01 1:08:32

Yeah, and I know even from my understanding from some of like the testing when you're even initially training data, like I know in several algorithms, like something deployed on the West Coast, they apply to patients in Florida and different outputs, different results, different scanner types, different population, you know. So I think that's something that they've just realized how you just need more and more data to accumulate. But you're saying, you know, just because something's FDA clear doesn't mean it's good to go, right?

SPEAKER_00 1:08:57

No, I mean, you know, it it's and there there have been very smart uh position papers put together by some of our society guilds, ACR, RNA, uh, and some other guilds, put a put a joint one together. Uh and uh it it does take monitoring, perhaps just as diligent as FDA clearance. I I dare say that the the on the ground post-FDA uh data, um, let's take Rad Partners who has you know 4,000 radiologists and 50 million exams a year or something like that, that that is a lot of data points to use to uh validate an AI tool, far greater than the number of data points that were used to initially get that tool approved. So uh we need to we need to do this as a specialty together. I mean, when when when a tool is released uh, you know, in in radiology at large, those data points need to be constantly monitored for performance and drift. And and if we don't do maximum effort there, we're not gonna have transparency and trust from patients, uh, and just as important from radiologists who are gonna need to trust this stuff. I mean, if you had a self-driving car, would you just trust the brochure that you can go to sleep and let it you know drive on the highway? I mean, you're gonna you're gonna need to see some real convincing and continuous monitoring of of uh those promises. Yeah, absolutely.

SPEAKER_01 1:10:17

So, Dr. Harvey, this has been so eye-opening for me. You know, as someone in like the more academic radiology sphere, we dived deep into the reading room and radiology AI. Um, and I was able to actually see, you know, what the newest AI looks like on the larger scale. You know, you come from a very unique practice where you guys have 4,000 radiologists across the country and you're working for this corporation that is able to adopt these at scale and test these at scale. So I think you have probably the most like up-to-date version of what is actually happening now at the most kind of like the pinnacle of AI. And it's really eye-opening for me to hear about this, and I'm sure our audience loved it too. So I just wanted to thank you so much for coming in. Um, really appreciate it.

SPEAKER_00 1:11:02

Thanks a million. Yeah, you know, I I'm glad we have this conversation. Everybody talks about the headwinds in radiology. We need a tailwind, uh, need to catch a break, and and this kind of technology is a great tailwind. I really enjoyed talking to you about it.

SPEAKER_01 1:11:14

Yeah, thanks.