Radiology Can't Keep Up. Here's Where AI Actually Helps | Dr. Nina Kottler

Speaker A: Every year, the amount of imaging is an order of magnitude that's 10x bigger than the amount of capacity that's being generated every single year. Now, in the past, we always had more and more imaging, but we had cycles. This is not a cycle. This is like a direct line that starts to go up like this with a 10x order of magnitude difference in the next 10 years.

Speaker B: Welcome to Frame By Frame Rethink Imaging, a podcast by Imlogix. 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 Rethink Imaging. My guest today is Dr. Nina Cutler. Nina is the chief medical officer at Mosaic, the technology arm of Radiology Partners, which is the largest radiology practice in the country. She's a practicing radiologist with more than 20 years reading cases. And she came into medicine from a rather unusual direction with a graduate degree in applied mathematics and optimization theory. Before she ever picked up a scan, she was the first radiologist to join Rad Partners. She spent years building and deploying clinical AI at a scale very few people have worked at. North of 20 million patient exams. She serves on AI Quality and Informatics Committees across the ACR, SIM, RSNA, and back in 2018, she won the Trailblazer Award for her work in imaging informatics. I really love talking with Dr. Cutler today. She's a brilliant mind and truly at the forefront of the field. Honestly, I barely referenced any of my prepared questions for Dr. Cutler as the conversation was just able to flow. She has such, like, a comfort and an ease about her, all while being such an incredible educator. I think y' all will really enjoy this episode. We touch on a lot of different points of clinical AI applications. We get into human trust, we get into data validation, we get into roi. Take a listen and enjoy the show. Welcome back to Rethink Imaging. I am so, so thrilled to be joined today by Dr. Nina Cutler. Nina, welcome to the show.

Speaker A: Thanks for having me, Chris. Excited to be here.

Speaker B: Hell yeah. I already have some intro questions for you and I'm already going off course, but I've been think I just came from AAPM where there are obviously a lot of conversations about AI and radiology, and I've just been thinking about this in general with like, the societal backlash to like, large language model and generative AI, and I just. Because we're going to be talking, uh, about AI and radiology, and I think a lot of people just have it a little twisted. What is AI. What's machine learning? What's generative AI? Where. What are the clinical applications? I was wondering if you could like, give us a bit of like, a system or like, what's the term? I'm looking for, like a classification of these different buckets? A little bit, yeah.

Speaker A: And it's funny because the backlash against AI is as it's getting bigger and bigger and more capable. It's using resources that people are afraid we're already running out of. But AI has been around for a really, really long time. I think it was first. The term machine learning first came out in the 1960s, believe it or not. So, uh, we've been using some AI, at least a version of AI and radiology, for years, since like the 80s. So going back, the original things that we called artificial intelligence were really machine learning. And you mentioned that term. Machine learning is just a way for a computer to try to figure out how to do things that we do today. And when we first started doing it, we would program it, we would tell it. Well, if you're looking at, in radiology, we look at a breast cancer. If we're looking for a breast cancer, it should have these 20 characteristics. So look for those characteristics. And so you're giving it the rules of what to do. Over time, we've evolved and that machine learning became more and more capable. It became deep learning. Deep learning is just a lot more richer data. The more data you have, the better it gets. And then now, today, what we're doing is we're moving to something called foundation models. And that's probably what everyone is comfortable with or what everyone's at least using today. Those foundation models tend to be large language models, at least that's what they started as. And large language models are language only. They're very large systems with lots of data, and they are using a different kind of architecture than the deep learning that we've used in the past. And that architecture enables them to be far more capable. And we are using those capabilities every day. They are now evolving into multimodal, multimodal language, multimodal models. Multimodal foundation models are now language, vision, video, all kinds of other components that are combined together. And if anyone is using ChatGPT, or Gemini, Gro, Claude, any of those, they're multimodal AI. They are also generative AI. Now, generative AI is something different. It's not just automatically a part of what we're doing. Generative AI means when the AI is generating an output that's not necessarily a deterministic Output. It's an output based on how the AI has learned. And that output could be a language output. Just like when we're chatting with a chatbot, it's giving a language output. So that's a generative AI language output, but it could also be a vision output where we're asking it to make an image or a movie. And, and so the, all the terms are related, but they're not all exactly the same. In radiology specifically, we're using two kinds of foundation models, the things that we talk about today, large language models and vision language models. Because in radiology we use a lot

Speaker B: of imaging M. And so. And you're primarily focused on, on like clinical applications of these models, Correct?

Speaker A: Correct. How do you them? Um, in a way in healthcare that we could add more value back into the system, which means improve the quality, decrease the cost, decrease the burnout, and add m. More capacity.

Speaker B: Yeah. And so for a while it seemed like as these models were growing, growing and learning, they were trying to not just improve their findings, but like find everything, see everything, catching what people might miss. But you've said at this point in the timeline, that's not really the problem anymore. And so I'm curious where things are in that journey.

Speaker A: Yes, we are evolving. We started seeing this new version of AI, which was not the programmatic AI, where we said look for these findings and if you find it, then you'll tell me. But where it learns on its own, either through labeling or through a lot of data. We started doing that back in healthcare in 2016. And in 2016, AI came out in radiology. That's why we have 10 years of experience with it in this specialty. That's why we talk a lot about what's happening, because we're a little bit ahead of everywhere else. Now, back in 2016, the major problem was what you said it was quality. How do we improve the quality? Because there are a lot of findings that we miss. Or maybe over call. Humans are not perfect. And especially as we get busier and busier, that imperfection is more obvious and it has an effect on patients. And that is important. Right. We all care about that. So the goal was how could you improve the quality of the radiologist by detecting more findings. That was the biggest one. It's, uh, enhancing the sensitivity of the radiologist without over calling. And a lot of the AI tools did that back then and still do that today. Now, why am I saying that it's different today? Because it's not. Because quality isn't important. Quality is Always the most important thing, but frankly the worst quality exam is an exam that you never get to. And that's the problem of today. The difference between 2016 and 2026. In those 10 years we have overwhelmed the system with so much imaging and not enough humans, radiologists to read that imaging. That capacity, volume, mismatch has caused massive amount of problems. Massive. Like turnaround times are getting skyrocketed. What is the turnaround time? It's how long does it take for the study to be interpreted. You could go to an outpatient imaging center for a cancer follow up and not get your results for weeks or months even. I've heard some places that are nine weeks behind. That's just crazy. And what about you go to the er? You know, the er, they're overwhelmed, they're ordering lots of studies. We used to have to have all of our imaging, our stat imaging done for the within 30 minutes on average. But it used to be closer to like 16 minutes on average. It was fast. Like uh, you get the study ordered, it would get imaged and then you know, within 16 minutes or so you would get a result. That's great. You could bring people through the ER quickly. It's not measured in that uh, mechanism anymore. It's measured now maybe in 45 minutes on average, an hour longer. And with that, patients are just sitting and backing up in the ed. It's costing hospitals a huge amount of money. Patients are getting frustrated and if you're delaying their care, have an effect on their health. And something that people might not actually know that's happening is because it is so expensive to hire radiologists right now because there's not enough of them. So everyone's in a food fight for the radiologist. So salaries are going up. Reimbursement for radiology has been going down for years. So expensive to hire, reimbursement coming down, massive mismatch. Like all of a sudden there's a lot of radiology practices that just can't survive anymore. They can't manage their own volume because they can't hire people. So they're going to hospitals and hospitals are paying subsidies to the radiology practices just to keep them in business. Millions of dollars. So capacity is a massive problem. It's a problem for patients, it's a problem for healthcare systems, it's a problem for radiologists. And radiology is a specialty.

Speaker B: I saw that you just were speaking with Ian Weissman. We had him on the show a little while ago where we were talking about this. I mean particularly within the framework of staffing. And we didn't get as much into like the volume side of the story as well. But it's such a, it's such like an interesting moment, right? Because like, yes, of course it's horrible that people are backed up and that we need more eyes on these images. But like, as myself and as we have been thinking about this rise in volumes, like, I'm so curious about the story behind it, right? Like a lot of this started happening around Covid and there is this downside to it, but at the same time, like, there's theoretically a lot of good coming from the fact that these volumes are going up as so many different people are getting more and more access to imaging. And so it's like, it's an interesting little moment, right? Because there is, there is so much like we're in the dredges, we're overwhelmed. But at the same time, like, I am hoping that the rise in volumes is indicative of a higher quality of care. Maybe, maybe not.

Speaker A: I wish it would be indicative of a higher quality of care if the right imaging happened for the right patient at the right time. We had new imaging that did new things. I actually think a lot of it is because we have patients that are getting older. Patients in the US are very sick. We're not a super healthy population. And with that, uh, there's more people going through the system and more imaging happens. And imaging is very important in those people that just never needed it before. 85% of the time when a patient has, uh, a problem and they go to a clinician, 85% of the time they're getting an imaging study to help figure out what's going on. So yes, that part, absolutely. Imaging is one of the best diagnostics that we have. I think though some of the imaging being done is not necessarily as important to do. Some of it is because we just need to get patients through the ED faster. We see a lot of studies that are being ordered not by the physician, but just based on the triage or there's more nurse practitioners and advanced practice providers that haven't had all of the training or experience that some of the other physicians have. And they might. We find that they order a bit more exams. Now, is that the primary problem? I don't think so. I think that imaging is useful. You're right. I think the population is getting older and more and more imaging studies are being done. And with that it's not, we're not going to solve it by either cutting down the amount of imaging, even though that's part of a problem. We're not going to solve it by that, frankly. We're not going to solve it by increasing our workforce. Because the. I don't know if you know the numbers at all, because the numbers tell a really important story. Every year, the amount of imaging is an order of magnitude that's 10x bigger than the amount of capacity that's being generated every single year. Now, in the past, we always had more and more imaging, but we had cycles. Like every four years we'd have too few rads, and then too many rads and too few rads, and we go up and down. This is not a cycle. This is like a direct line that started to go up like this with a 10x order of magnitude difference in the next 10 years. Because it's supposed to continue. We are expected that we would need another 15,000 radiologists and we only have 35,000 in the industry to begin with. It's not possible to manage. You have to manage this with technology.

Speaker B: Yeah. Okay, so let's dive into it because. So when we talked before we recorded this episode, you were talking about your early days using these clinical AI tools. And you had four different machines running, like a sea of monitors, each of them running different AI tools. None of them connected. Can you tell? Can you just, like, paint us, like, a quick little picture of what that version of your workflow in daily life was like? Uh, and then give us, like, you know, we'll. We'll move into where things are now.

Speaker A: Yeah, I mean, where things are now. We're mostly still like that for most organizations. Who was it that said, the future is here, it's just not evenly distributed. Like, that's exactly who we are.

Speaker B: Right.

Speaker A: Like, that was a great quote and it's totally true for healthcare. So there's. In our organization, we've invested a lot in AI, so we're a very early adopter. So what I tell you we're doing is way ahead of the industry. But let me give you the baseline of where the industry stands. So remember, you kind of when you were giving that description, it reminded me of the movie the Matrix, where the guy's like, looking at and there's like green numbers. Like, that's kind of what it's like. Because we have multiple different systems in healthcare that we use. We use an electronic medical record which has some information in radiology, we use our PAC system, which shows the images and some of the data. We've got a risk, we've got all of these different information systems. That don't connect to each other or if they connect is very superficially so on. When I open up a study, I have four monitors. I do. I have four monitors and they each have multiple different applications on them. And guess what? The person or the thing that has to consolidate all that information is a human. So I look at the EMR for some things, I look in the wrist or something else and then I take it all together and then I concentrate on the images. But guess what humans are really bad at doing. Humans are terrible at integrating information. This is horrible. Like, it's just not what we were meant to do. But, uh, computer systems are great at it. It's just that we've never had a system that overlies everything that can translate all of the information. And many of our systems were built in the 1980s, 1990s. EMR is a little bit later, but that makes it very difficult. So that standard of care today is crazy. Different amount of systems, the clinicians sitting in front of it, integrating all that information. Where are we moving to? What we want is we need all that information. It's not that it's not important, but we want it to be presented in a way, kind of like the movies, like Iron man, he's wearing his helmet and there's tons of data points it could give, right? But it gives us stuff he needs at that moment based on what he's looking at and right what he's going to be doing. And that's the kind of system we need. So what we're moving to is away from these individual AI tools that provide you with one detection output, to tools that help with the workflow. Integrating information together, summarizing it, providing it where you're, so that you as the radiologist can continue looking, looking only at the images because that's the patient and that's what matters, right?

Speaker B: And so, yeah, and it feels so, once again, very novice here. But it feels like a lot of the data that you're talking about, right? Like stuff from the emr, all that data, it's a lot of it. It's relatively easy to find and acknowledge that it's like reached ground truth, right? Like looking at things like dose or image quality or like there are a lot of numbers in there that you can say, like, yes, this is right, even looking at the, like summaries and stuff. But when looking at images, I feel like that's where things like, get a little bit murkier. So I'm just curious, like, what does the ideal workflow look like in Your mind and like, just. Just talk me through it a bit more, I guess. Yeah, sorry. This question's a little disjointed. I acknowledge that.

Speaker A: Well, let's answer part of it at a time and then we'll answer the other, uh, part. So part of it, it sounds like what you're speaking about is AI can hallucinate. And how do you know when the answer it's giving you is right? Sometimes it's really easy to verify because it's a number that you could just look somewhere else and verify that number. Other times it's something that's making a decision about something. And how do you validate that? Because we know AI isn't perfect. And the goal of the outcome that is going to be the best outcome is when you can combine the human and the AI together in a way that they're both better. If you have one that works really well, the other works really well, but they don't work well together, then. Well, frankly, you're actually not helping the patient because these AI systems are not autonomous, and autonomous means they work fully by themselves without a human in the loop. Almost every single one of these systems are meant to be an adjunct. So you need the two to be working better together. And actually imaging mostly is an easy thing to verify, not always. So if we get right now, let's just talk about the NAR AI solutions that are out there. The Narai solutions are solutions that will look at images, maybe even the priors, and it will tell you if there's something on that image that the AI sees. So it might say there's blood in the brain. It's kind of what Google used to do when you would bring up a picture of a cat and it would say cat. Then you bring up a picture of a dog, and actually Google would say dog for the image of the dog. But our imaging AI system will say not cat. That's the, that's the level we're at right now. For these nary eyes to say there is is blood in the brain. I see it on this head ct, or I don't see blood in the brain. And then another tool will say, there is a hole in the lung, or there is not a hole in the lung, there is cancer. There's not cancer. So ours are all like binary? Yes. No, for these.

Speaker B: Oh, interesting.

Speaker A: I, uh, know it's very different than people probably realize because it's completely different than what you're using day to day.

Speaker B: Yeah.

Speaker A: So that's kind of state of the art right now. For most people using AI and those tools, they are additive on top of your PAC system. So how do you validate them? Generally, it's pretty easy. If an AI says there's blood in the brain, I look for where the blood in the brain is. Now, it's very helpful if it tells me a little bit more than just saying there's blood in the brain, because there's hundreds of images on a CT scan of the brain, and you have to look through all of them, and it can be really subtle. So if it shows you a picture of where it sees the blood in the brain, or if it tells you an imaging number where it sees the blood in the brain, great. I can then look and say, you know, it's either there or it's not there. What's hard is when there's more quantitative imaging, and there is some of this in AI right now, where it will look at the brain and it will say, the volume of this part of the brain is X and the volume of that part of the brain is Y. Well, we don't measure volumes. And so how do I validate if that volume is right? That's when it gets a little more difficult. So, yes, some things are very easy to validate, some things harder. And we need to be able to make sure that the human and the AI system work really well together. So we have to be thoughtful about how do you put them together in a way that the human can validate the AI because the human is ultimately the one responsible. Right. Tools are not responsible for humans. Humans are responsible for humans. So you need the clinician to be armed with the things from the AI decisions it's making. Transparency about how it's working in order for them to be able to agree when the AI is right and disagree when the AI is wrong. Not always obvious.

Speaker B: Yeah, I'm curious about this. Like, yes, no binary approach to tools.

Speaker A: Amen.

Speaker B: How long. So is this at, uh, Rad Partners that y' all are doing this, or is this kind of everybody there?

Speaker A: It's everyone else, and outside of language. So anything with computer vision has to go through the fda. And you cannot not sell that tool until it is authorized by the fda. And the FDA has not authorized any of these generative AI tools that can kind of do anything yet. I think that will happen next year. Which means that everyone is still using for computer vision. Everyone is still using these narrow AI tools because they have been authorized, they've been cleared by the FDA. So those narrow tools are. You could have 30 of them you could have a hundred of them, each doing different things, but you, you know, you need quite a few of them. I would say most people probably have five or 10 at most. So they're only looking for five or 10 different things in your images. And, uh, where we're going is to a place where the AI will look at everything in your head CT or everything in your brain mri, and it could dictate an entire report. That's what we're doing right now at, uh, Radiology Partners. And it is very different than what is happening in the world. And it's because we are clinicians. We created this tool. And when you create the tool, you can use the tool under medical exemption, which is if you're a medical practitioner, you should be able to use and try tools in order to make the system better, because you need clinicians to make the system better. We're also rolling it out under a research protocol. So that enables us to do something that the industry is not yet doing. That's why there's a difference.

Speaker B: Yeah. And with that, that by definition, would that be agentic AI? Uh, if we're continuing to come back to buckets.

Speaker A: So let's, let's talk about that because I think everyone uses that term incorrectly and. Yeah.

Speaker B: Myself included.

Speaker A: Well, you know, it's because everyone is saying, uh, everyone's using it as a hype term. Like if you say, I'm using generative AI and I'm using multimodal AI and I'm using agentic AI, well, maybe people will like, invest in your tool. So everyone's calling it. But let's define that a little bit better so it's clear. Remember how in the very beginning you asked me, like, what was machine learning? And I said, machine learning are some of the earliest versions of what we considered AI was us programmatically telling the computer what to look for. And in that kind of scenario, we are making up all the rules. And in healthcare, if you can imagine, and that would mean, if you wanted to do that, well, you had to very. You had to predict every scenario that occurs if you want to make a rule that will work no matter what situation you're in. In healthcare, there's just too many edge cases that you can't predict every action that's going to occur. So people started talking about agentic AI. Well, why? Because agentic AI can be more predictive and adaptive. So let's define an agent or an agentic AI. An agentic AI system has to be able to do a few things. Number One, it has to be able to take a high level goal and break that down into rules that it would create that it will use apply to be able to get toward that goal. So for me, if I'm uh, I'm going to give you a radiology example because this is how we're thinking about using it. But right now, in order for me to determine what study do I read next, I have a system that has a bunch of rules that are built in and it says, well, if it's a stat, if it's from the er, then do that first or if it's an outpatient, maybe do that a little bit later. It's rules based. What if I change that to an agentic system? Well, I would basically tell that system I wouldn't give it rules, no rules. I would say I'd give it a goal. I want you to create a work list that gives me the highest quality output and it meets the best turnaround time so we get it done the quickest. Right. That's all I would give it then. So it has roles then. Two, you have to give it access to different parts of your system. You have to give it access to different components of data. Well, it needs to know what imaging studies are coming in, needs to know what the schedule is and what radiologists are on. It might need to know what subspecialty the study is. It might want to figure out what the quality of the study is. So you give it access to a bunch of tools that give it information because it can't possibly do anything without data or information. So it then breaks down and creates its own rules. Once it creates its own rules, it has to be able to monitor itself. So number one, high level goals. Number two, access to things so that it can get to those goals by creating its own rules. Number three, it needs to be able to self m monitor. As it self monitors, it realizes, am I going away from my goal or am I going toward my goal? If it's going away from a goal, then it automatically changes its own rules. It adapts its rules. Doesn't require me to come in and say, no, change your rules. Adapts the rules itself to move closer and closer to the goal. That's what an agentic system is. And if it doesn't do those three things, I would not call it agentic. So agentic is not a, uh, better AI, it's just using an LLM. Agentic, ah, is something that is enabled within your system.

Speaker B: Got it, Okay. I very much appreciate that. And also the self Monitoring, Yeah. Is the piece of the puzzle that I was definitely not queued into as like a general rule, because, like, that, I mean, that the m. The monitoring for AI drift, it seems to be like a pretty crucial component just in. In terms of like all. All sorts of different models. Right. Like outside of the realm of radiology as well.

Speaker A: Now, AI drift is a really good topic to talk about because the name itself is actually not what it is. It makes us think that the AI model itself is drifting. The model does not drift. What happens is the data coming into the model is changing. You mentioned Covid before. So before COVID happened, we didn't have a lot of this exact kind of COVID Right. We didn't have a lot of disease that looked like Covid does. So if you had trained an AI model to look for pneumonia and then all of a sudden you had Covid cases, would it identify it well, or would it not? Might not, because it hasn't been trained on that before. So in general, what we say is a model is either generalizable or not generalizable. What makes it not generalizable is the training data. If you trained it on a small subset of data, then it will do really well on that subset of data. But if you give it something different, it doesn't necessarily do well. A generalizable model is hopefully trained on a much larger data set. And so it works well across a bigger both patient population, but also exam population, maybe a different manufacturer of the CT scanner or a different protocol. All of those things are very important. And it's another benefit of M moving toward these much bigger foundation models because they're trained on so much more data than these narrow AI systems that we're using today.

Speaker B: I get the scale of the data, but especially like as you are trying to build these models, how do you monitor the quality of the data itself? Right. Is the concept that a certain amount of scale will start. Start to directly improve quality? Are you trying to meet scale and quality? Like, what is. What does that balance look like?

Speaker A: I think scale does help with quality. So the more data that we give, the better it tends to be as long as your data is good data. Now, right now, the difference between a narrow AI system and a foundation model is how you train it. So these narrow AI systems, you train it by giving it label data. Well, what is label data? It means that a human would go in and they would look at the image. Maybe it's a chest X ray. And if that narrow AI was looking for pneumonia, they would like, literally Circle the pixels that would have pneumonia, or they'd show on the image. They would write, this image has pneumonia and this image doesn't have pneumonia. That's a human labeling the data. As you do that, the data itself becomes structured because I am providing structure, and that's easier for a computer to understand. So you don't need as much of the it. The difference between that and what we're moving toward. We're moving toward these foundation models that just are trained with a massive amount of data, millions of images. We can't possibly label all of them. We don't have enough rads to do the regular work, let alone the labeling work.

Speaker B: Right, right. Yeah, that's what I'm getting at.

Speaker A: It's weird, right? So. So you would think that, well, gosh, if I don't have someone labeling it, maybe it's going to be worse. It ends up that the more data we give it, the AI is actually pretty good. If you give it the report, the radiology report and the images, the AI is pretty good at figuring out where the pneumonia is on its own. Now, the key is you want to look and make sure that that data doesn't have really bad results in it. So not every human is created equal. Not every radiologist is equally as good at each system. And there may be some radiologists that they aren't as good at reading certain study types. And you want to be able to look at your data and exclude those or the ones that are really good at reading it. You want to make sure you're elevating them, because in theory, if you just took a bunch of reports and you created an AI system off of that, you'd think the AI is going to be as good as the average radiologist who created that report. But you want your AI system to be better than the average rad. So you have to go through your data and just siphon through to make sure you're optimizing your data so that your model output is optimized. And what we found is you do that, uh, there hasn't been, like, we haven't noticed the limit, that if you give it more and more data, it can still get better. You know, it's kind of perhaps asymptotic, and there might be a point at which it's not worth it because it costs a lot of money and to train these, and no one's paying for these systems, so you have to be able to make it manageable. But, um, the more data you give it, in general, the better.

Speaker B: Yeah, we've touched a lot on the data and the data science of everything. But I do want to touch a little bit on the more human element. Talking about, like, the trust in these systems and the human trust in these systems and the patient trust in these systems. Because it's a relatively high barrier. Right? Like, yes, there is societal distrust of AI because of environmental reasons, because of bandwagon reasons, because of like, artistic integrity, like, uh, you know, every. Everybody's got their own boat. All valid, sure. But as like a potential patient, I'm kind of curious, like, how, how do you see the process of rolling these out to the public and just kind of public perception in general?

Speaker A: Yeah, I think a lot of people in the public are probably using these tools already for themselves, asking questions to the AI because it's. You get an immediate answer, you get a lot of data. It's very, very personal to your information. Now, I don't know if people are sharing their. I've seen certainly some people sharing their data with these public systems and, um, other people that are worried about that. I would say in general, people like having knowledge because knowledge enables us to do something. And if you want to take control over your own health system, having that knowledge about your own health is important. Important. Now when we talk about deploying AI within the system, so not a patient using it outside the system. So if we're creating it within a healthcare system, there are a lot of regulations and a lot of safety requirements that we have to manage through. And so privacy is extremely important. And it's been a very big public debate. How do you make sure that patient health information is safe and not being used against them? We've had those privacy rules for a really long time in healthcare. So we're. I think we're pretty good there. I think where the danger is, is yes, if you want to share your information with a open system, then in theory that your information is then out there. But if you're doing it within something that is deployed by a hospital or FDA authorized, then we are following the guidelines that are already required and have been in place for a really long time. So I'm less concerned about patient privacy. My, my thought is that AI has so much potential value in healthcare in particular, because it's one of the only things that can solve this capacity problem, enhance the quality of what we're doing. I think it's gonna help transition us from. Right now we're in a place where we are trying to treat a population that's called population health metrics. I think we're gonna start to go from there to precision medicine, meaning deciding what's wrong for that patient in particular, maybe even for that. Their genome, their genetics, their labs, all of that's going to start being combined and be much more important for that actual patient. And we're moving to a place where instead of determining and diagnosing what already exists, it's going to be about predicting what's going to come. And that's much more impactful. These are the kinds of things that AI can do and that's so important. And I think within the system that we're already in and we have enough guidelines that are protecting patients that I see a massive, massive amount of benefit of moving forward with these. And I hope that patients understand that because maybe they think about, well, what's happening outside of healthcare, and what's happening outside of healthcare has a different set of rules than within healthcare.

Speaker B: What you said about like, people craving knowledge is really, really resonated in this context because especially when you're talking about predictive findings, right? Or findings, whatever you want to call them. Because, like, I, I think that there is something like, so human and tempting, right, about being able to predict and know what could or will happen. Right? I mean, you saw it with, with all of the like, genetic cheek swab companies, 23andMe, and all of those different DNA tests. You know, people are just so curious and excited to learn about themselves that they were like, a lot of folks were just like standing off their DNA, not reading the fine print and just saying, like, hey, as long as I know that I am, you know, predisposed for this condition, like, I'm happy. It's a really interesting point. So. But like, what about, what about like a skeptical radiologist who's being told they need to change their workflow?

Speaker A: Yeah, that's hard.

Speaker B: Yeah,

Speaker A: it's hard because change is scary. No human likes change and change in an environment where the environment is extremely stressful right now, we're completely overwhelmed with volume and there's a lot of legal liability. We're responsible for patients and patients lives. Like when you start to make a change with something, it adds a degree of uncertainty and it adds extra work that we're not. That we just generally didn't have to do. And I'm gonna give you an analogy. You know that humans cannot do two things at once. Despite what they say, we actually can't. Right? Like, we can only do one thing that requires our frontal lobe executive functioning at the same time. So why can we ride a bike and think about what we're going to say on a phone call. We can do those two things at once. Why can we breathe and drive and do something else? That's because the second thing is built into our basal ganglia. That's a part of our brain that enables us to do functions without thinking, without using our frontal lobes. Of the way that a radiologist functions in this super stressful environment is all those clicks and the integration of how do you pull information from one place to another and how do you scroll through an exam and where do you look? That has been ingrained. We've been doing it for years. I've been in practice 20 years. You get really good at that. And it's kind of like riding a bike. If all of a sudden you then say, yeah, I'm going to turn your handlebars now so that when you turn to the right, you're actually going to the left, all of a sudden you have to think a lot about what buttons to push and how to do that. And that means it's harder to do some of the. And that's scary because ultimately we don't want to harm patient care. And so any change that we institute is a hard thing. Now, does it mean we shouldn't do it? No, absolutely not. We need to continue changing and evolving. You just need to do it in a safe way. And everyone is going to have a different amount of risk aversion. Some people are going to be really excited about the new technology and they're going to want to adopt it right away, and others are not. And so it's about how do you just manage change across that spectrum to make sure the people that are really excited about it aren't using it inappropriately? And also the people that are scared of it start thinking about using it in the right way because it will make them better. That's the whole point. So it is a lot of education, that's probably the most important thing. And some of it is about transparency. How do you make sure that the AI is giving the information that the radiologist needs in order to understand how it's making its decisions? And a lot of that is not necessarily happening from the vendors today. We have to tell the vendors we need information to make the human and the radiologist better together. Other. And that should be part of the regulation and part of the tool that is required.

Speaker B: Yeah, Obviously different types of healthcare facilities are going to have different levels of access to these tools. Right. Like, I mean, we, we see it with scanners, we see it with staffing, right? There are smaller rural facilities, big academic centers, you know, so if, like, what would you say to, you know, smaller facilities that are still doing like a decent amount of volume, they're still underwater, they don't have access to these tools. What sort of advice would you give to them in this kind of interim period? And especially as we are trying to safeguard, you know, the mental health and the burnout of our current rad professionals, in order to not make our staffing job even more? What, what sort of advice or direction would you give them? Kind of in this time, as things are ramping up, we need to be

Speaker A: not afraid of this. We need to be jumping in and using it now. And the hospital systems, the pract, any of the groups that are not are going to fall behind and the groups that are using AI will outperform them. So they may be thinking about it in the short term, like, oh, I can wait, I can wait. But you know what, that's going to hurt them in the long term. And so jump in, get involved. The people that get involved early are the ones that get to make decisions about where these things go. And it's really important for clinicians and healthcare systems to be making those decisions. Not, not just vendors per se. It has to be a group of people. Now the hard part is there's no return on investment from a standpoint of cms, Medicare, private payers in general are not paying for these tools. So how do you afford them? How do the smaller hospitals do this? And each AI tool has to have its own return on investment. Depending on what the return on investment is, tells who should pay for it. So if the return on investment is, is radiologist becomes much more efficient, then sure, the radiology radiologist should pay for that. If the return on investment is more patients go through the system, they get better downstream care. There's savings for the hospital because patients get through the system faster. Better experience than the hospital should pay for that. And if the AI tool doesn't have a return on investment, it's dead in the water. So it has to have a return on investment. The thing that I want to point out though is there's something that we need to be doing that we call AI governance. And AI governance means how do you oversee the AI? It's kind of a question you were asking before to make sure it's safe and safe over time, and how do you make sure the end user is using it in the right way? That's a clinical governance. And not everyone can necessarily have the Dollars to be able to afford that. And this is where I've been speaking to people in the government and societies and I think should start talking to CMS about not reimbursing the AI itself. What if we don't. Right. What if the AI itself has its own return on investment, but instead we reimburse the clinical governance that is overseeing these tools to make sure that they're deployed appropriately and monitored over time? And that would be my request.

Speaker B: Okay, interesting. And I have to follow up. How do you think about quantifying ROI in, in this context? Right. Like, obviously there's some cases where, you know, you could say, okay, we've done this many more exams since implementation. Like, you know, that's roi. But I feel like so many of these clinical tools, the ROI is kind of soft and nebulous. So how do you, how do you think about that?

Speaker A: You have to define exactly what the ROI is because the people that are buying it will involve the CFO and the CEO and not the C of a hospital who are generally buying these tools. And they are going to care about the exact dollars and where they line up. And so you have to break it down. It's not, oh, in general, there's going to be a return on investment because your patients are going to love you. No, it's about, you'll get this many more people through the emergency department and you are going to have this many more admissions. It sounds like, kind of hate talking about this because it sounds like a business and this. And I mean, the hospitals have to stay in business. So we're talking about it. Even though what we care about as clinicians is patient care, but we're talking about the business side, capitalist society.

Speaker B: We get it.

Speaker A: Okay, thank you. All right.

Speaker B: Uh, we exist in it. It's part of the conversation.

Speaker A: It is. And the way that most hospitals are paid are what is called fee for service. They do more services, they get more fees. So if they do more procedures, and now, uh, we want them to do the right procedures, but let's say they do more procedures or they bring patients from one part of system into another. We have something called breast arterial calcification, which is you can identify calcifications when you do a mammogram and you're identifying these calcifications within the vessels of the breast ends up that, that has some prediction to if someone's going to have a risk of heart disease, of coronary artery disease. So what if you take those patients that are getting screened for breast cancer and you send them over to cardiologists because many women present with their first heart attack with no warning signs. Well, what if we screen them now? You're getting more patients into another part of the system, and those can add dollars. So you have to break down and think about how do you actually make sure that the dollar, the dollars, can recoup the cost so that we can put these great tools into place.

Speaker B: Well, as much as I hate to end with capitalism, this. This is. This is about all the time that we have today as. Although I could keep talking to you for hours.

Speaker A: Hours.

Speaker B: Dr. Kotler, thank you so, so much for coming on the show and for talking to me today. It's been amazing having you.

Speaker A: You're very welcome. Um, I appreciate it, Chris. It was. It was a very fun time.

Speaker B: Thank you so much. Frame by Frame Rethink Imaging is brought to you by imlogix. 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 imlogix is rethinking imaging in healthcare, visit imlogix. 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 Imalogix, thanks for tuning in.