Rethink Imaging

EP 24 • July 30, 2025

Radiology AI Reality Check: Automated Reports, Implementation Failures and Agentic Future

Featured Guest

Dr. Woojin Kim, MD

Chief Strategy Officer & Chief Medical Information Officer, HOPPR; Chief Medical Officer, ACR Data Science Institute • HOPPR

Listen Now47 min

Also on:

Let's Talk

Beyond dose

Ask us what your imaging data can actually show you.

Start a Conversation

Share This Episode

Subscribe to Rethink Imaging

Host Chris St. John talks with Dr. Woojin Kim, Chief Strategy Officer and Chief Medical Information Officer at HOPPR and Chief Medical Officer at the ACR Data Science Institute, about where generative AI in radiology actually stands. At his ACR 2024 Morton Lecture, Kim predicted automated draft reporting would reach CT exams, and within a year models like MedGemini, Medversa, and Merlin proved him right. A prospective Northwestern study measured a 15.5% documentation efficiency gain from AI draft reports for radiographs. Yet adoption stalls for a human reason: radiologists like the way they sound, and some attendings delete technically correct drafts and redictate from scratch.

The second half is a candid look at what stands between promise and practice. A fabricated finding in a medical report carries direct patient consequences, and Kim is blunt that general-purpose models like ChatGPT are not ready for imaging diagnosis: they are not medical devices, and confident misreporting far outnumbers the correct cases shared online. No FDA-cleared device today uses a large language model, and liability between radiologist, hospital, and vendor remains unsettled. Kim closes with his vision for agentic AI, drawn from a radiologist who read three to four times normal volume with six scribes, and his standing advice: learn about AI, learn to learn, think critically, and get your hands dirty.

Host

Chris St. John

Host, Rethink Imaging • Imalogix

Featured Guest

Dr. Woojin Kim, MD

Chief Strategy Officer & Chief Medical Information Officer, HOPPR; Chief Medical Officer, ACR Data Science Institute • HOPPR

Watch the Episode

Key Takeaways

Automated draft reporting moved from prediction to publication fast. After Kim called it at his ACR 2024 Morton Lecture, models including MedGemini (Google), Medversa (Harvard), and Merlin (Stanford) appeared, and a prospective Northwestern study showed a 15.5% documentation efficiency gain for AI-drafted radiograph reports.
Personalization decides adoption. Some attendings select all, delete, and redictate technically correct resident drafts, and models tuned for academic benchmarks like BLEU scores have almost no correlation with what matters clinically.
Hallucination risk has improved but not disappeared, and the rules lag behind: zero FDA-cleared devices use large language models today, and liability among the radiologist, the hospital that bought the tool, and the vendor that built it is unresolved.
Agentic AI should work like the six scribes Kim once watched support a radiologist reading three to four times normal volume: agents running background tasks, connected to clinical tools through the Model Context Protocol.
Kim’s advice for imaging professionals: learn about AI, learn to learn, think critically, and get hands-on with the tools instead of only reading about them.

Full Transcript

[00:00:01] Woojin Kim: I’m going to get controversial here, but let me tell you that not only will you see AI automatically generate draft reports for chest X rays, you will see automated draft reporting in CT exams. Now to be honest with you, for chest X rays, it wasn’t much of a prediction because I knew, technologically, we already had the capability. For example, at the RSNA two thousand twenty three annual meeting, when I was a CMO at Bet dot ai, we demonstrated a proof of concept for automatically generating a report from a chest X-ray. Now the prediction part was talking about doing the same thing with CT exams. So was I right? Most recently, a paper from Northwestern prospectively evaluated a generative AI model capable of providing giraffe radiology reports for radiographs showing 15.5% document efficiency gain.

[00:00:54] Chris St John: Welcome to Frame by Frame Rethink Imaging, a podcast by Imalogix. Here, we explore the intricate world of medical imaging, aiming to dissect the field and inspire both professionals and curious minds alike. I’m your host, Chris St. John. Welcome back to Frame by Frame Rethink Imaging. Today, we are thrilled to be joined by doctor Woo Jin Kim, a renowned radiologist and pioneer in the field of medical AI With over two decades of experience in clinical radiology and health care innovation, doctor Kim has become a leader in integrating artificial intelligence into medical imaging. Currently, he is the chief strategy officer and chief medical information officer at Hopper and the chief medical officer at the American College of Radiology Data Science Institute. Doctor Kim’s work spans from cofounding successful health care startups to developing AI tools that revolutionize diagnostic workflows. Today in this episode, we’re gonna explore some of the challenges and some of the potential of GenAI in radiology and the importance of balancing, once again, human expertise with technology. Welcome to Frame by Frame, doctor Kim. It’s great to have you.

[00:02:08] Woojin Kim: Thank you for having me on, and thank you for that kind introduction.

[00:02:12] Chris St John: Yeah. Absolutely. So I have a bunch of questions prepared. Before we get into those, let’s just, like, start high level. One, can you just tell me a little bit about yourself and how you found yourself, honestly, on this podcast and doing what you’re doing? Sure.

[00:02:30] Woojin Kim: Yeah. I’ll be more than happy to share my background. So, again, my name is Woojin Kim. I’m a radiologist by background, and my specialty is in musculoskeletal radiology. And, you know, I did all my training at UPenn starting from medical school, residency, and even MSK fellowship. But during the year that I did my MSK fellowship, I took, you know, one day out of the week. So every Monday through Thursday, I would do MSK fellowship. But every Friday, I would take Amtrak from Philadelphia to Baltimore to do a second fellowship in imaging informatics. And it was during that year that I learned a lot about imaging informatics, and it was also the year that I started my first company in radiology search engines. Then I decided, you know, you know, after I finished my training, you know, I have no interest in academic radiology. I’m gonna go into private practice. So I actually went into teleradiology for about two years. But about, you know, after two years, I decided, you know what? I think I, you know, I want to do something different. So despite my original, you know, decision, I decided to go back to UPenn to be a faculty member there, and I, you know, climbed the, you know, the classic academic ladder, eventually, you know, becoming, you know, interim section chief and everything. But also during that time, you know, when I became an attending physician, I started my second company, Montage HealthCare. And as the second company ended up buying out the first company, so the first company exited that way. And Montage, we helped to grow for about six and a half years, and we were installed in over 500 facilities, both US and Canada. And that was a search and data mining applications for radiology, and we were acquired by Nuance. And I was CMIO at Nuance for about three and a half years, and I was there until, you know, before they got acquired by Microsoft. And I left Nuance to start my third company called Equivalent Intelligence that was using AI. I was trying to bring something that was used in supply chain world, like demand forecasting and capacity planning, and try to bring that into radiology. And I grew that company for about two and a half years, and we got acquired by Red AI. And so I became the CMI or head Red AI for about two years. And then after our CNA last year, doctor Kon Siddiqui, and this all connects because he was one of the mentors that taught me imaging informatics twenty years ago. And he used to joke around saying, hey. You know what, buddy? If you ever start startup number four and I’m not part of it, I am never talking to you again. So I said, hey, buddy. I I I I really need you to help, you know, help her out. And to buddies of mine, you know, John Pilat and, you know, Bill Boone from Red AI, we decided, you know, hey. After thinking about it, hey. You know what? Red AI is growing rapidly. It’s in a very, very good space. And, you know, we felt that pretty comfortable leaving Red AI where it is and then to join us, you know, much smarter, smaller startup to help that grow. Because if you look at my history, you can tell that, you know, I like to take something small and and help it to grow. And so we you know, during the two years that we were at Red AI, we really helped to grow the company significantly. Obviously, right now, it is one of the hottest AI company in radiology. And so we felt comfortable leaving that company and then help out another company, Hopper. And so I’ve been with Hopper since January this year. And then the other hat that I wear is I’m also the chief medical officer for the American College of Radiology Data Science Institute. So that means that, you know, I also have to pay attention to the policy side and, you know, aspect of the AI in radiology. So those are some of the things I do. I still practice radiology very little, but I do. And then I guess my unofficial hat, full time job, is giving talks. So somehow I guess I got labeled as one of the Gen AI guys in radiology. And so I’ve been invited to give a lot of talks. And I think this year, I’ll probably end up giving about 45 talks at this rate. So that’s another thing that’s keeping me motivated.

[00:06:28] Chris St John: Lot of presentations.

[00:06:29] Woojin Kim: Okay. Pause right there. Yeah.

[00:06:31] Chris St John: I love it. I mean, honestly, it’s an interesting story. Like, uh, it’s fun to see you bouncing around and just, like, swapping out all of these hats. Yeah. Congratulations.

[00:06:42] Woojin Kim: Thank you. And it’s been fun. I’m very blessed.

[00:06:45] Chris St John: Well, good. Okay. Well, like, let’s get into it. So I did some research. We have some questions prepared. I did my research with the help of GenAI. I was saying this to doctor Kim before we started the recording, but for the authenticity of the experience, I can say this to my listeners as well. But getting into it, at at ACR twenty twenty four, doctor Kim, you predicted that we should expect a fully automated reporting system for chest X-ray and CT. What makes you feel confident that that level of automation is coming, and what needs to happen to implement it safely in a real department?

[00:07:25] Woojin Kim: You’re right. At the ACR twenty twenty four annual meeting last year, I was given the opportunity to deliver the Wharton lecture, and there I said, you know, I’m going to get controversial here, but let me tell you that not only will you see AI automatically generate draft reports for chest X rays, you will see automated draft reporting in CT exams. Now to be honest with you, for chest X rays, it wasn’t much of a prediction because I knew, technologically, we already had the capability. For example, at the RSNA two thousand twenty three annual meeting, when I was a CMO at Bet dot ai, we demonstrated a proof of concept for automatically generating a report from a chest X-ray. Now the prediction part was talking about doing the same thing with CT exams. So was I right? Well, later that year, several papers were published that feature that particular technology, including MedGemini from Google, MedVersa from Harvard, and Merlin from Stanford, to name a few. Now interestingly, the latter two became start ups. Most recently, a paper from Northwestern prospectively evaluated a generative AI model capable of providing draft radiology reports for radiographs showing 15.5% document efficiency gain. However, like you asked, there are many things that need to happen to implement this safely. First, you know, obviously, it needs to generate accurate reports. Second, it needs to be tightly integrated into the radiologist workflow and incorporate personalization. Also, one aspect of radiology workflow the developers need to remember is what I call the dimension of time, and I don’t mean turnaround times. We radiologists compare the, you know, the current exam with prior exams and reports, And if these AI models cannot do that and only look at time point zero, well, that has limited clinical utility. Now after that prediction, I also said, by the way, if I said this five years ago, I would have been booed out of the room. However, you know, when I said that no, I also knew that I wouldn’t be booed out of the off the stage. Why? Because I’ve noticed a shift in perception regarding this. So for example, at RSNA two thousand twenty three, on the first day of the conference, guess what? I showed that particular capability, that proof of concept to only a very few people that I knew personally because I wasn’t sure what they would say. You know? I didn’t want them to say, hey, Woojin. What are you working on? Are you trying to replace us? You know? But I can tell you by the next day, we were showing it to everyone because they were all saying, hey. When can I have this? So the increasing burnout and radiology shortage, which I’m sure, you know, you’ve spoken with other, you know, uh

[00:10:11] Chris St John: We’ve covered that in quite a few episodes.

[00:10:14] Woojin Kim: Times. Right. And that had changed things, and that’s why I felt pretty comfortable making that prediction as well as that statement in front of a thousand radiologists.

[00:10:23] Chris St John: Yeah. For sure. And then, you know, RSNA twenty twenty four, that was my first time attending, but it was I couldn’t take two steps without hitting another AI product, AI presentation talk. You know, it’s it’s around for sure. Honestly, like, candidly, even to the point where I you know, only being around for a couple of years now myself, I’m like, yeah. I’m we are talking about it. Like, it’s hard to not have these conversations, I guess, is what I would be saying.

[00:10:54] Woojin Kim: You’re right.

[00:10:55] Chris St John: You have said that technology should work for you, not the other way around in the context of AI. Can you give us an example of a GenAI tool that maybe looks great on paper, but in practice kind of, I don’t know, fails the usefulness test? And what are the developers or researchers missing about radiology workflows?

[00:11:18] Woojin Kim: Well, why don’t I keep using the same example to illustrate the potential weakness? So, for example, you know, the creating an automated draft radiology report from images has great potential uses. Now if you follow me on LinkedIn, and you will see sometimes I will make posts where you see a post with lot of red circles in writing, and you know that there are a lot of issues with that particular paper. And oftentimes, it’s because the researchers and developers fail to incorporate clinical domain expertise. And if I don’t see a single radiologist on the author’s list on a radiology AI paper Mhmm. For me personally, that’s a huge red flag.

[00:11:57] Chris St John: Yeah. Absolutely.

[00:11:57] Woojin Kim: Not always, but oftentimes, it causes lots of issues. Unfortunately, you know, these generated outputs are often optimized for academic benchmarks like, you know, BLEU scores, but those metrics have almost no correlation with what actually matters clinically. Now I’ve seen many, you know, quote unquote state of the art models on paper that confidently tell you, you know, and I see that on the figures and why that’s why I circle in red. There is an interval, you know, increase since the last exam when there was no prior exam or findings that simply aren’t there. And speaking of priors, I always look for prior exams and read the prior reports when they exist. You know, one of my attendings always used to tell me, nothing makes you smarter than the prior reports, so don’t forget to look at them. So if the AI solution only looks at the current exam, well, that’s not going to be very useful. And, unfortunately, right now, a lot of the solutions only look at the current exam. In fact, the dimension of time is one dimension that is often missing from these research papers, and the another major challenge will be the editing cost. So for certain exams, I will tell you, it’d probably be much faster for me to dictate from scratch than to, you know, try to fix mostly correct AI report because, you know, that means I have to reverify every single line to make sure it’s not hallucinating something critical. So if you’re developing these solutions, you have to ask yourself, due to their design, am I increasing or am I decreasing the cognitive load of the radiologists? And finally, if your AI solution, generative AI or not, isn’t integrated into my existing workflow, I can tell you I don’t care how pretty your application looks. That’s an example of what I what you asked me earlier about, you know, technology not working for me, but the other way around. So the bottom line is that, you know, the designers and developers need to start in the high pressure reality of an actual radiology reading room. They need to go there and sit there and watch the radiologist work rather than, you know, chasing leaderboard scores in isolation, which a lot of papers do.

[00:14:02] Chris St John: Yeah. I mean and while we’re talking about reports, right, I mean, radiologists are extremely particular about the report style. You’ve noted that adoption improves when the AI sounds like a radiologist is using it. Can you just build on that a little bit more? Like, why is it so important, and how is it impacting patient care and communication rather than just personal preference?

[00:14:27] Woojin Kim: Yeah. And, you know, I say this often in my talks, and I usually say, hey. I can say this because I’m a radiologist. We like the way we sound, and I say this. And I can say that because, you know, I’m a radiologist. Right? The caveat, however, is that everyone is different, and so let me give you an example to illustrate what I mean by this. Let’s say a resident or fellow, you know, drafts a report about, you know, a particular exam after a readout session with an attending radiologist. Afterward, that attending radiologist has to go into the reporting solution and, you know, sign off on that report. On one end of the spectrum, you have attendings who will be like, you know, yeah, sure. It doesn’t sound like me, but it’s good enough, and they’ll just sign off with very little or no editing. However, I’ve also seen the other side of the spectrum, and this is real, where literally the attending would go, select all, delete, and redictate the whole thing from scratch even though the draft report is technically correct. Why? Because it doesn’t sound anything like them. Now if you speak with a lot of the practice leaders and department heads, they will give you they’ll tell you where, you know, they like to ideally have more uniform standard reports. However, there are also plenty of radiologists that can tell you who simply won’t use a system if it generates a text that doesn’t sound like them. The post editing work isn’t worth the time. And this is why for many AI applications, I advocate keeping in mind the importance of personalization. Now it’s not necessary for every solution in the radiology workflow, but it is a component often overlooked in research papers and many commercial solutions as well, and yet it’s an important aspect I personally observed in my industry role. So I’ll give you a real example. So when I was at Red AI, you know, obviously, you know, it’s a it’s a company that’s been known for creating the automated, you know, impression generation. And it if you hear why they like it, one of the reasons why radiologists like this particular solution is they will tell you, like, yeah. This thing sounds like me. And, you know, in Einstein, they actually did an experiment one time where they, you know, printed out a whole bunch of reports that were generated by the radiologists and then whole you know, print out a whole bunch of reports that were generated using this, you know, automated solution. Impression generation, they mixed them up and they showed the radiologist, hey, which one was, you know, radii solution versus you, and they couldn’t tell the difference. And so for me, when I heard that, I realized, like, yeah, I have a intuitive sense that personalization is important. But when I see it in the actual, you know, clinical use, then I really got to, you know, get a new appreciation for the importance of this. Not everyone’s gonna care about personalization, but I’ll tell you for those who do, that actually makes a huge difference. And that’s one of those things that I rarely ever see in research papers. They’re always trying to, you know, have some kind of a metric and say, hey. We just beat this, you know, metric by point zero five, you know, type of thing. But at the end of the day, just, you know, we’re human, and you gotta have that human aspect.

[00:17:33] Chris St John: That’s exactly what I was about to say. Like, it is just so deeply inherently human to want that from the content. Right? Like, even just when I’m, you you know, using ChatGPT to help, like, clean up an email or whatever it is, as soon as I see that, like, stereotypical AI voice tone and just, like, loaded with EM dashes, I’m just, like, I lose my mind. Like, it it like, it drives me insane up the wall. I’m like, no. No. No. No. No. No. I like, this feels inhuman, and therefore, I do not like it.

[00:18:10] Woojin Kim: Perfect. Perfect. I think people now that they’re using ChatGPT, they have better appreciation of exactly what I’m talking about. Right. And it’s funny that you’ve mentioned the em dashes because I I saw that, you know, many posts people saying, oh, yeah. I know when you use ChatGPT GPT because you use m dashes. The thing is I’ve been using m dashes long before Chat GPT. It’s really, really I feel like, oh, man. Like, I don’t wanna delete it because I don’t wanna sound like Chat GPT. But but here’s the thing. That to that point about the automated impression, the reason you know, that was actually the thing that became really popular right after Chat GpT came out from a research perspective because people realized, like, oh, I can use, you know, chatbots like Chat GpT and LLMs to automatically generate impression. Now I have to say, Red AI has been doing this commercially long before Chat GpT, but the thing is that, you know, people started publishing papers on this and say, hey. Yeah. You can use Chat GpT, and, you know, it can generate, you know, impressions for you. But I’ll tell you, what I say in the in the in my talks is that, yeah, if you will give her a ChatChippy Tea, a set of findings, and ask you to create an impression, it does create a pretty good looking impression. But to exactly what you said, the reason why you personalized your Chachapiti instance, I’m sure you have done that with prompting and examples and projects and all kinds of tools that we have, is because, hey. You know what? That doesn’t sound like me. I’m not gonna copy and paste that thing in. And same thing happens in radiology too. It’s like, yes. Technically, it’s correct, but I don’t sound like that. I don’t talk like that. And, therefore, if I were to copy and paste, I’m gonna have to edit it. And as soon as I edit it, I lose all my efficiency gains, and it’s just better for me to do it on my own. And so you really need to think about a personalization. And, thankfully, things like ChatGPT, because you’re right, now people are pretty attuned to it. Like, oh god. This guy use a, you know, chatbot for his, you know, writing. So you were trying to, you know, make it, you know, help you write like you.

[00:20:01] Chris St John: Mhmm.

[00:20:02] Woojin Kim: And the personalization, I think people have a, you know, a renewed appreciation for that aspect of AI tools. Yeah.

[00:20:09] Chris St John: Absolutely. I just small little anecdote. We I just got sent a link from my employer for, like, you know, a company workspace within ChatGPT, and I had been working on, you know, my version of it for the last however many months. And now having to start fresh again. I’m like, no. So I’m, like, having it generate all of these prompts to, like, try and transfer my voice over, but immediately within the new workspace, I was just like, nope. Doing it myself. Just like this is, like, trying to reteach it. I’m like, I don’t have the time for this. I’m just gonna edit I’m just gonna edit this on my own. Exactly. Okay. And so, I mean, let’s pivot a little bit. You already mentioned hallucinations a little bit ago, and it’s something that you know, it’s inevitably gonna come up when talking about GenAI. In radiology specifically, though, how big of a risk is the AI making something up? Have you seen examples of, like, confident misreporting, and how do we mitigate that risk for these reports?

[00:21:12] Woojin Kim: Well, hallucination or confabulation is absolutely a legitimate concern in radiology AI, and, frankly, it should be. And unlike, you know, Chepa’s making out facts about movies Right. When an AI fabricates findings in a medical report, that directly impacts patient care. And I’m sure, you know, you’ve seen plenty of ex post where, you know, people uploaded their images to Chat g p t or Grok and shared their outputs. But let me tell you, as of today, these general purpose AI models are not ready for making any type of imaging diagnosis no matter who says what. First of all, they’re not medical devices, but more importantly, for every correct case that you see online, I can tell you there are many, many more alarming examples of confident misreporting and misdiagnosis. And I just recently read, like, I think earlier this week that, you know, these chatbots have gradually you you know, companies have removed the standard disclaimer about how these should not be used for medical diagnosis. And I’m sorry, but that is simply irresponsible.

[00:22:14] Chris St John: That’s been it’s been taken out?

[00:22:16] Woojin Kim: So it used to be always when you put in something about medical diagnosis, you know, related thing, first of all, sometimes, you know, these chapa say, hey. Look. I’m not I can’t do this. And then but, you know, people have found a way to, you know, get around it. And if you do make them do it, it will always say, hey. You should consult, you know, medical professionals. You know, this is not a real medical diagnosis, blah blah blah. So they would have the standard disclaimer, but this, uh, research paper just came out. I it hasn’t gone through the peer review process yet, but they have noticed that there’s been a significant decrease in the number of times that these things will actually show you this disclaimer. I’m sure it’s hidden in some fine print somewhere that no one reads, but, like, actual interactions with a user, it just doesn’t show. And and, you know, people have you know, the authors have made some, you know, some, you know, hypothesis as of why this is happening, but the reality is it looks like they’re doing it less and less and less. And I think, you know, this is dangerous. I mean, yes. I mean, people probably will ignore those disclaimers anyway, but still, like, as a company, you can’t just, like, not say it because you think people are gonna, you know, ignore it. So but the the scariest part about this hallucination or confabulation is how plausible these reports sound. You know? You know, they use, like, proper medical terminology, and they follow standard formatting. I mean, if you didn’t look at the images and you just looked at the outputs, they look completely legitimate even to a radiologist. And so this is why, you know, I always emphasize we need human in the loop workflows where radiologists maintain, you know, full control. This means, you you know, instead of simply saying AI alone is better than AI plus human, which is another topic. By the way, there are a lot of issues with these, you know, types of papers from a methodology perspective, but we need to figure out at the end how to create that optimal human AI symbiosis. For example, you know, the AI should show its work, highlighting, you know, which parts of the image support each finding it reports, and the radiologists themselves also need to be educated not only on AI and its limitations, but also on the best ways to interact with it. Kinda like the way you describe, like, over time, you learn how to use ChatChippity and these tools to, you know, really help you do what you do. So I advise in my lectures, typically, first, learn about AI. And especially for younger audience, I say, number two, learn to learn. And third, learn to think critically. And I think that last part is, you know, something that we’re losing. So I would say approach AI as a copilot, not autopilot.

[00:24:53] Chris St John: Yeah. I mean, absolutely. And it it to me, it feels like right now, like, GenAI is it’s almost like the new methodology of how we interact with the Internet, which is wild. But, like, you know, I’m I I rarely go to Google anymore to search for something because why would I? You know, it’s I’m getting fed a bunch of ads first, and then it’s just like a mess. SEO is confusing and confounding, and I can never find what I want. When I can talk to my little chatbot friend, and they pull up exactly what I need, you know, it’s like it’s the whole methodology of how we are interacting with our technology is changing into a conversation, which to me is just really fascinating.

[00:25:41] Woojin Kim: It is.

[00:25:42] Chris St John: So, you know, you you made this transition from reading scans to building AI tools. How has the shift changed the way that you view the daily work of a radiologist?

[00:25:58] Woojin Kim: I mean, you know, to be fair, you know, I still, you know, practice radiology. And so for me, it’s it’s always the same thing, like, comes down to the same thing, which is to really try to identify the pain points and try to solve that. Even though I have switched, you know, mostly to the industry side, I do still practice, granted, not every day. And people ask me, you know, what do you why why do you still practice radiologists? And there are a couple reasons. You know? One of them is I know radiology, but I also need to feel the pain that the radiologists are going through. Knowing about them is not enough. I can read about all the challenges of radiologists, but it’s very different when I’m actually in the reading room and going through the thing and realizing like, hey. You know what? This is really painful. And that helps me to be a better developer and better, you know, entrepreneur. That’s one thing. And a little bit more selfishly, I will tell you, you know, if I’m in a room in front of a bunch of radiologists and I’m trying to sell my solution, it’s much more effective for me to say, hey. I’m a practicing radiologist than to say, hey. I used to be one. And so

[00:27:09] Chris St John: For sure.

[00:27:10] Woojin Kim: There’s some street credibility, you know, some street withered as well too.

[00:27:13] Chris St John: We can’t mock it.

[00:27:15] Woojin Kim: That’s why, like, for me, it’s, like, I I do, you know, try to, you know, still, you know, maintain a you know, keep my foot in the the trenches a little bit because I I need to feel the pain, not just know about the pain.

[00:27:27] Chris St John: Right. And so, I mean, you know, you you have your entrepreneur hat. Right? And then you also have your ACR leadership hat. You know, as you wear you know, to throw both hats on, if you could, what do you think the biggest barrier to implementing Gen AI in radiology is right now? Is it technical? Is it cultural? Is it legal? Is it you know?

[00:27:47] Woojin Kim: You know, that’s a great question that really gets to the heart of what we’re dealing with in 2025. And the short answer is it’s multifactorial. Right. Yeah. You know, the hallucination and confabulation problem that we just described is improving, but let’s face it. It still has not disappeared. So since we’re just specifically talking about GenAI, given the gravity of the problem that, you know, these things can cause, it is still an issue that you cannot ignore, especially in medicine. And the other big barriers are, like you said, regulatory and liability clarity or, frankly, the lack thereof. You know? For example, we have zero FDA cleared devices that use large language models today, and there is a lot of uncertainty and lack of clarity on this. And the liability question is another one. Yes. The radiologists take liability at the end, but, you know, what about the hospital that purchased it or the company that made the AI solution? And right now, there is a lot of uncertainty. And, you know, I sometimes get asked questions about this too, about the medical you know, the legal liability about all these solutions. And my simple answer is, you know, honestly, I don’t think we will know until there is, like, a massive lawsuit, a class action lawsuit about this. And then depending on how the judge and the jury and they decide, we’ll determine, like, yeah, this is what the medical liability looks like. Because Mhmm. Until we I mean, I don’t wanna see that either, especially, you know, wearing the hats that I wear. But I almost feel like something like that has to happen for us to get clarity on it because right now, there is no clarity. It’s everyone’s guess. Like, if you ever go to a panel discussion about you know, if you post that exact same question right now, people will say, well, I don’t know because there is no FDA clearance, you know, devices that use as a lab. So it’s anybody’s guess. You know? Anybody’s, you know, opinion is just matter of opinion. And same thing with the medical legal liability. How much liability do these things carry? No one knows because there’s no clarity on it. So these are some of the challenges, but I suspect, you know, more and more as, you know, these solutions get used, I think we’ll hopefully get some clarity, but it is impacting some of the decision making and development processes and things like that. So, yes, good question.

[00:29:55] Chris St John: To take it even further, right, I saw that you were you’ve been discussing I don’t even know if I’m pronouncing this right. Agentic AI? Yeah. Agentic. Yep. Agentic. Yeah. Yeah. So for folks who are new to the term outside of me who just read a GenAI definition of it, Would you mind describing Agentic AI in specifically, like, within the context of radiology and what it could actually do?

[00:30:22] Woojin Kim: Yeah. That’s a great question. And, you know, you’re right. And most people are just like, man, I just got used to understanding what Gen AI is, and then everybody

[00:30:30] Chris St John: Next level.

[00:30:31] Woojin Kim: A Genentech AI, AI agent. So let me first talk about AI agents. And you’re right. There has been considerable hype around the term AI agents or a Genentech AI lately. In fact, many have referred to 2025 as the year of AI agents. Unfortunately, though, despite widespread discussion on this topic, there is no universally agreed upon definition, so I cannot give you that. But I think it’s easier for me to explain it by describing what it isn’t. So let’s say, you know, you ask an LLM based Chapa like Chesapeake to give you a day trip itinerary to San Diego just because I live in Southern California, and it gives you this great itinerary. Now when you ask which day of the week next week you have your day off, guess what? It won’t be able to answer because it doesn’t have access to your calendar. Often, these tapas don’t have access to proprietary knowledge and tend to be passive, meaning they won’t do anything until you ask a question. Okay? So keep that in mind. So, you know, what if you connect that chatbot to your calendar, which you can do now? Sure. You know? Then you will be able to answer that question. But what if you follow that question with another question saying, hey. Okay. Great. So can you tell me the weather for next Wednesday? And it will fail unless that chat also has access to local weather data as well. And, sure, you know, you can connect it to a weather service API, but this is simply what I call an AI workflow. I can connect many additional services and tools, but at the end, if a human is the decision maker, there is no AI agent involvement. In contrast, when you give an AI agent a goal, it will proceed to reason, typically using an LLM, to plan and act, meaning, you know, it will execute using tools, and some people, you know, call this combination react framework, you know, reason and act. And now some people, you know, consider, though, this is enough to call something an AI agent. If it can reason, enact. But for me, I think it needs to then go further to observe together feedback and reflect, meaning, you know, review and adjust strategies, and finally, iterate, which I think is a really important component to me personally about what constitutes an AI agent, meaning that it will repeat as needed to ultimately achieve a goal. So you give a go, AI agent, you know, keeps iterating, comes up with a plan, iterate until it achieves a goal. And another, you know, simple way to think about it is to think of an AI workflow as being more linear. You know? Whereas an AI AI agent or agentic workflow is more circular. And, of course, you know, not to get even more complicated, not to confuse even more, but believe it or not, some people actually even differentiate the term agentic AI, which is a framework for turning a goal into plans Mhmm. From AI agent, the thing that actually performs the action. So some people use it synonymously, so and some people are like, no. You can use them synonymously. They’re two different things. So but you will see these two terms in the media quite a bit. So one thing I would definitely tell you, probably, I’ll be honest with you. I don’t usually like to make predictions, but if I were to force to make one about Agenstic AI in radiology is that at this, hey. I’m sure you’re gonna see vendors claiming to leverage AgenTic AI, when in reality, they’re using, like, if and then, deterministic AI workflows. Okay? Now to be totally clear, there’s absolutely nothing wrong with an AI workflow. In fact, I usually advocate from a more practical perspective, a hybrid approach of using AI workflow and agentic AI for many use cases, but just be wary of the industry hype around them because that industry hype that you’re seeing outside is gonna come to radiology and in our arsenal, you are going to see this. So since you asked about radiology specific example, so I’ll go into that a little bit now that I describe, you know, what is an AI agent. In radiology, GenTech AI is like having an intelligent colleague who can, you know, independently manage an entire workflow to achieve a broader goal. Right? So to reiterate, you know, what makes something, quote, unquote, a GenTech is is autonomy and adaptability. So what could it actually do? Well, you know, this goes back to my emphasis on the non interpretive use cases, some of the, you know, things that I focused on in my previous startups where, you know, instead of just, like, looking at the the imaging value chain in radiology where everybody just focusing on the image interpretation aspect of it, you can imagine having a GENTIG AI benefiting all aspect of the imaging value chain. You know, picture an AI that handles your entire preread workflow, you know, triaging urgent cases and selecting optimal imaging protocols and gathering all relevant patient data. And, you know, during interpretation, it’s not just, you know, flagging abnormalities, that’s what you’re seeing today, but performing, like, comprehensive analysis, you know, measuring tumor volumes. Then it knows that, oh my god. Now I need to track it, you know, changes over time and then realize that, okay, I need to then take this and integrate with the lab results and clinical notes to suggest differential diagnosis for you. Now to your question, you know, is this science fiction, or is this reality? And I would say it’s not a moonshot thing. Okay? We’re erasing some of the building blocks for this, believe it or not. So in the near future, I think you can expect augmented AI to handle sequences of tasks with human oversight. So just to highlight why this is not a science fiction, this sim this past sim this year, the Hopper team actually Gentic AI in radiology by using MCPs or model context protocols, another term that you probably heard quite a bit outside of radiology, to provide greater context to radiology during interpretation by using Fire MCP and also DICOM MCP.

[00:36:23] Chris St John: Would you mind explaining that term for me, please?

[00:36:26] Woojin Kim: Yeah. Model context protocol?

[00:36:27] Chris St John: Yeah.

[00:36:27] Woojin Kim: Yeah. Sure. So model context protocol, really, really simply put, you know, some people have, you know, described it almost like it’s not the perfect you know, most perfect description, but I think, you know, to keep it really, really simple, it’s like having USB C or creating it. So for example, like Okay. You know, when the AI agent I remember I told you that, you know, a agent would what makes it agent is that it’s ability to use tools. Right? And for me to use, like, tools, different tools, say, you know, for radiology, for me to say grab, you know, the DICOM images or go to the EMR and get the patient’s history or the reasons for exams or lab values. Okay? I need the agent to be able to use tools. But here’s the thing. From the agent to use tools, because these tools are made by different folks, they need to use different APIs.

[00:37:17] Chris St John: Right? Right.

[00:37:18] Woojin Kim: And there’s a lot of complexity around that. Right? And and as an agent builder, agent needs to, like, know which API to use and make sure that, you know, it’s, like, it’s deprecated. Oh my god. Why is it not working? Oh, they got a new version and blah blah blah. And this API works different than than this API and this API, so it becomes extremely complicated. The beautiful thing about the, uh, model context protocol or MCP is that it uses the same method to connect with all the tools. Kinda like instead of having USB a, c, and all of these things Yeah. Literally, I don’t care what you do. I just need to just use the same thing to connect with you. So it’s like being everybody’s saying, like, hey. We use USB c. So if you got a USB c cable, you can connect, connect, connect, connect, connect. It’s kinda like that to put it the most simplest way. I mean, obviously, there’s a lot more

[00:38:07] Chris St John: That’s great. Thank you.

[00:38:08] Woojin Kim: But the idea is the MCP allows it to allow you to, like, connect to a lot of different tools in a very easy manner. And so so I don’t know if you use cloud, because I know you mentioned, you know, HTTPT, but if you use cloud, you know, one of the things that I use all the time, and I absolutely love it, is I use the cloud desktop, and I use the MCP, and so now if you actually are, you know, using cloud, you will be able to see, like, how you can connect with different services, and so, you know, I connect with my Gmails and my Google Calendar all through the cloud interface, and I could literally ask questions, you know, saying, hey, you know, what meetings do I have today? And can you, you know, just like you could literally imagine if you were to be able to prompt the cloud and say and in fact, I do actually, I have I do this every morning. This is my daily routine. I have a, like, actual prompt where I say, hey. You know what? Go through my calendar and tell me who I’m meeting with. When was the last time I, you know, I had discussion with them? Because so go through my email chain and summarize, you know, my previous conversations with this person. Why am I meeting today? What’s the objective? What’s the goal? And then if, you know, it’s about a company, then I actually asked, you know, Claude to say, hey. Give me a summary of what this company does. And so it it would then give me a whole, you know, today’s itinerary, and it’s doing that through cloth, and it’s

[00:39:24] Chris St John: That’s very cool.

[00:39:25] Woojin Kim: Because of MCPs. And so what we did at SIEM was to, hey. Let’s bring that in radiology. We got all these disparate silo systems, and so let’s, by creating these, you know, fire MCPs and DICOM MCPs and be able to, you know, show this. And so we it was a hackathon prototype, but I think this is another exciting thing that a lot of people are not talking about it right now, but it’s something that I think, uh, you will see more and more because and and it’s going to, like, really explode the development as we are seeing outside of radiology with you know, you’re seeing new things coming out every single day, literally, and that’s because of, you know, the openness of the community and things like this. Yeah.

[00:40:08] Chris St John: Yeah. I mean, it yeah. It feels like the the point in the exponential curve where we’re, like, right before we go vertical.

[00:40:17] Woojin Kim: Yeah. But if you were to ask me, and I I I’ll share this with you because I shared this example illustration in my talks to talk about agentic AI and the future of it as it relates to radiology. I mean, I obviously shared with you what we did at the hackathon, but, you know, my personal agentic AI choice, if I were to able to do it today, it would be something like something I saw, you know, several years ago where I saw a teleradiologist. He said, hey. You know what? I I wanna show you how I work. So a number of colleagues and I, you know, went to visit his reading room, and when we went in, he had six huge monitors in front of him. Right? And in front of his desk, which was, like, really big, he has six mice, And that’s because there was no IT system that can handle this guy’s unique workflow I’m about to describe for you. So behind them, though, set six scribes with their own workstation. Alright? Now each of the six scribe was responsible for one of the six monitors. You see where I’m going with this. Right? And so when he sat down to work, he would look up, and the scribe number one would put up the first case. Scribe number two, second case, and so forth. So he will look at the first screen, and he says, it’s a chest X-ray, left laurolone pneumonia, left side of parofusion. That’s all he would say. Then he moves on to the second case, and he uses second mice and, you know, scroll up and down. He says, a small vessel ischemic disease impression, low cutie, and control cranial hemorrhage. That’s it. And he goes on and on. Right? And so by the time he gets to case number three or four, scribe number one says, hey, doctor so and so. I have first case ready for you. And then the scribe would put up the, you know, the the reporting solution screen with all the structured reports and the reports all made out, and then he just looks and goes, yeah. That looks good. Sign off. So here’s the thing. This guy was reading three to four times what most radiologists would read, but the difference was that he looked extremely relaxed, so relaxed that, you know, after three or four cases, he would turn around and, you know, he would chat with asking about us, and and then he goes to, you know, case number five and six and chatting with the scribes, and he was very relaxed. And so when I saw that, I realized, like, oh my god. Imagine. Like, you know, looking back, you know, this was several years ago. And for me, like, when you ask me about Adjenti AI in radiology, I’m thinking, like, what if instead of those six scribes, you had six agents, you know, just running in the background doing all that for you. How much would Radios would love that kinda solution? So that’s kinda, like, where I see as sort of the North Star for some of the not the only solution, obviously, but one of the potential solutions when it comes to GenTIGAI. And hopefully, that explains with that beginning to the end. We would now have a better understanding of, yeah, now I just I think understand agentic AI.

[00:42:55] Chris St John: I like that example that you gave me as well because upon your description of agentic AI, there’s there was almost a piece of me that sound like, I was like, it sounds like the human element is being removed to some degree, but it was nice to to to hear the larger context in which you’re talking about.

[00:43:13] Woojin Kim: That’s right. So I wanna emphasize that rather than replace radiologists, this technology has the potential to create a true partnership that can make us radiologists more efficient and accurate, and that’s for me, as as a radiologist, that’s how I want my agent to work, not like, oh, there’s a radiologist agent, and he or you know, it could just go and read a whole bunch of cases, you know, on its own. And that’s not that’s not, you know, what I envision as, quote, unquote, agentic AI. Yeah. Yeah. Great.

[00:43:41] Chris St John: Doctor Kim, thank you so much for joining us. Do you have any final thoughts to give our listeners before we let you go today?

[00:43:48] Woojin Kim: Yeah. So what I would say is this is an exciting technology, and I really urge anyone listening to learn about AI. You know, I said earlier, like, you know, learn about AI, learn to learn, and learn to think, you know, critically. All these things are important. But, you know, when I say learn about AI, it’s not just about reading about it. And I’ll be honest. You know, I’ll tell you. I’ve been very fortunate, and I’ve been very blessed, and I’m extremely grateful that every company that I’ve been with, whether it’s mine or, you know, uh, somebody after being, you know, acquired, they always allow me to, you know, keep reading. And I spend about four to six hours reading about AI every single day. And I know that most people don’t have that kinda luxury and time, but, you know, one thing I will tell you is that it’s not just about reading. You really have to play with the AI solutions. You gotta actually get your hands dirty. And so one thing that I recommend is instead of just, like, reading about AI, that’s a great start. But I really encourage people to play with different tools, like, you know, whether it’s ChatCPT or Cloud, Midjourney to make images or v o three to, you know, create some videos or, you know, hey, Jen, to, you know, create an avatar of yourself to, you know, just wear 11 laps to just to see, you know, what, you know, AI voice’s technology is like. Because you get to learn about its limitations and what it can and cannot do by actually using AI. So one thing I highly suggest folks is to really use AI and not just, you know, read about it. And this is going to be extremely important. And so one thing I will definitely close out by saying, you know, AI should be absolutely be our, you know, powerful assistant, you know, helping us catch things we might miss, improving our efficiency, even, you know, suggesting differential diagnosis. But at the end of the day, a human radiologist needs to be the one, you know, that’s making the final call and standing behind it. And so one of the things that if I can really encourage the developers that are listening is to really think about the domain expertise. This is not just about technology. It’s not just about the algorithm or the model architecture. There’s a real human behind this technology that’s using it to impact other human other patients’ lives. And so, you know, you gotta look beyond just the technical aspect of it, really incorporate clinical domain expertise. And instead of trying to, you know, replace radiologists, think about, like, hey. How can I augment the radiologists in this process? And so really think about how do I, you know, improve the human AI symbiosis instead of simply saying, oh, obviously, AI is doing it, you know, better by itself. At the end of the day, when you get sick, you are going to find a human doctor, and you want that human doctor who sees you to be at the best at, you know, his or her, you know, practice. So so at their you know, be able to practice at the top of their license. And so if you can help them do that, that’d be great. So those are some of the things that I like to, you know, conclude by. Yeah.

[00:46:57] Chris St John: Well, thank you so much, doctor Kim. It has been truly a pleasure having you today. Thank you so much for joining us on Frame by Frame, and we’ll talk to you soon.

[00:47:05] Woojin Kim: Alright. Thank you so much.

[00:47:08] Chris St John: Frame by Frame Rethink Imaging is brought to you by Imologix. Here, you’ll find engaging interviews with thought leaders, experts, and patients sharing stories that showcase the transformative power of medical imaging. To discover how Imologix is rethinking imaging in health care, visit imologix.com. Be sure to subscribe to Frame by Frame Rethink Imaging on Apple Podcasts, Spotify, or wherever you listen. And from all of us here at Emalogics, thanks for tuning in.

Questions from This Episode

Is AI already writing radiology reports?
Draft reporting for chest X-rays exists, and CT is following. Kim demonstrated a chest X-ray report proof of concept at RSNA 2020, and models such as MedGemini, Medversa, and Merlin extended the approach to more complex exams. A prospective Northwestern study found a 15.5% documentation efficiency gain for AI-drafted radiograph reports, with the radiologist still making the final call.
Can general-purpose chatbots like ChatGPT diagnose medical images?
No. Kim is direct that these models are not medical devices and are not ready for any type of imaging diagnosis, whatever viral posts suggest. For every correct case shared online, there are many more examples of confident misreporting and misdiagnosis.
What does agentic AI look like in radiology?
Kim explains it through a story: a radiologist reading three to four times normal volume, relaxed, because six scribes handled the background work. Agentic AI swaps those scribes for software agents running tasks behind the scenes, connected to clinical tools through the Model Context Protocol. There is no agreed definition of an AI agent yet, which is why he defines it by what it is not.