Rethink Imaging

EP 52 • September 3, 2026

AI Workflows, System Capacity, and Clinical Governance with Dr. Nina Kottler

Featured Guest

Dr. Nina Kottler, MD, MS, FSIIM, FAIM

Associate Chief Medical AI Officer, Mosaic Clinical Technologies • Radiology Partners

Listen Now45 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

Radiology practices face an expanding gap between imaging volume and available radiologists. Annual growth in imaging studies outpaces workforce capacity growth by ten to one, increasing turnaround times and creating operational backlogs across outpatient centers and emergency departments. Dr. Nina Kottler, a practicing radiologist and clinical technology leader at Radiology Partners, joins host Chris St. John to discuss how technical infrastructure can address these workload pressures.

The conversation covers the evolution of imaging algorithms from narrow detection tools to multimodal foundation models and agentic systems. Dr. Kottler outlines the three defining requirements of agentic AI, explains how changing patient data affects model generalizability, and addresses the financial realities of technology adoption. She then makes the case for formal clinical governance frameworks and for reimbursing technology oversight instead of individual software applications.

Host

Chris St. John

Host, Rethink Imaging • Imalogix

Featured Guest

Dr. Nina Kottler, MD, MS, FSIIM, FAIM

Associate Chief Medical AI Officer, Mosaic Clinical Technologies • Radiology Partners

Watch the Episode

Key Takeaways

Growth in imaging study volume outpaces radiologist workforce capacity by ten to one, shifting the main focus of radiology technology from detection sensitivity to operational throughput.
Agentic AI requires three distinct capabilities: translating high-level goals into self-generated rules, accessing data across connected systems, and monitoring performance so the system can adapt its rules over time.
Model performance depends on data generalizability. Changes in incoming patient data can reduce performance when algorithms are trained on narrow datasets.
Combining language and vision architectures allows systems to generate draft reports for entire examinations under research protocols, moving beyond single-finding detection tools.
Healthcare systems must evaluate return on investment against specific operational goals and establish clinical governance frameworks to oversee software deployment safely.

Full Transcript

Rethink Imaging Podcast Transcript

Guest: Nina Kottler, MD, MS, FSIIM, FAIM

Host: Chris St. John

CHRIS ST. JOHN 00:00:23 Welcome back to Rethink Imaging. I am thrilled to be joined today by Dr. Nina Kottler. Nina, welcome to the show.

NINA KOTTLER 00:02:07 Thanks for having me, Chris. Excited to be here.

CHRIS ST. JOHN 00:02:10 Hell yeah. I already have some intro questions for you, and I’m already going off course. But I just came from AAPM, where there are obviously a lot of conversations about AI and radiology. I’ve been thinking about this in general with the societal backlash to large language models and generative AI. Because we’re going to be talking about AI in radiology and I think a lot of people have it a little twisted—what is AI? What’s machine learning? What’s generative AI? What are the clinical applications? I was wondering if you could give us a bit of a classification of these different buckets.

NINA KOTTLER 00:02:53 Yeah, it’s funny because the backlash against AI grows as it gets bigger and more capable, using resources that people are afraid we’re already running out of. But AI has been around for a really long time. I think the term machine learning first came out in the 1960s, believe it or not. So we’ve been using some version of AI in radiology for years, since the ’80s.

Going back, the original things that we called artificial intelligence were really machine learning. Machine learning is just a way for a computer to try to figure out how to do things that we do today. When we first started doing it, we would program it. We would tell it, if you’re looking for breast cancer in radiology, it should have these 20 characteristics, so look for those characteristics. You’re giving it the rules of what to do.

Over time, we’ve evolved, and machine learning became more capable—it became deep learning. Deep learning is just a lot richer data. The more data you have, the better it gets.

Now today, we’re moving to something called foundation models. That’s probably what everyone is comfortable with or at least using today. Those foundation models tend to be large language models, at least that’s what they started as. Large language models are language-only. They’re very large systems with lots of data, using a different kind of architecture than the deep learning we’ve used in the past, which enables them to be far more capable. We are using those capabilities every day.

They are now evolving into multimodal foundation models: language, vision, video, and all kinds of other components combined together. If anyone is using ChatGPT, Gemini, Grok, or Claude, those are multimodal AI. They are also generative AI.

Generative AI is something different; it’s not automatically a part of what we’re doing. Generative AI means the AI is generating an output that’s not necessarily a deterministic output. It’s an output based on how the AI has learned. That output could be a language output, just like when we’re chatting with a chatbot, or it could be a vision output where we’re asking it to make an image or a movie. All these terms are related, but they’re not all exactly the same. In radiology specifically, we’re using two kinds of foundation models that we talk about today: large language models and vision-language models, because in radiology we use a lot of imaging.

CHRIS ST. JOHN 00:05:50 And you’re primarily focused on clinical applications of these models, correct?

NINA KOTTLER 00:05:57 Correct. How do you use them in healthcare in a way that adds more value back into the system—which means improve quality, decrease cost, decrease burnout, and add more capacity?

CHRIS ST. JOHN 00:06:09 For a while, it seemed like as these models were growing and learning, they were trying to not just improve their findings, but 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. I’m curious where things are in that journey.

NINA KOTTLER 00:06:33 Yes, we are evolving. We started seeing this new version of AI back in healthcare in 2016—not the programmatic AI where we said, “Look for these findings and tell me if you find them,” but where it learns on its own, either through labeling or through a lot of data. In 2016, AI came out in radiology, which is why we have 10 years of experience with it in this specialty. We talk a lot about what’s happening because we’re a little bit ahead of everywhere else.

Back in 2016, the major problem was quality: how do we improve quality? There are a lot of findings that we miss or overcall. Humans are not perfect, and especially as we get busier, that imperfection is more obvious and affects patients. We all care about that. So the goal was to improve the quality of the radiologist by detecting more findings—enhancing the sensitivity of the radiologist without overcalling. A lot of AI tools did that back then and still do today.

Why am I saying it’s different today? 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. That’s the problem today. The difference between 2016 and 2026 is that in those 10 years, we have overwhelmed the system with so much imaging and not enough radiologists to read that imaging. This capacity-volume mismatch has caused a massive amount of problems. Turnaround times have skyrocketed.

What is turnaround time? It’s how long it takes 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 even months. I’ve heard some places are nine weeks behind. That’s crazy.

Or say you go to the ER. They’re overwhelmed and ordering lots of studies. We used to have all of our stat imaging done for the ER within 30 minutes on average—closer to 16 minutes on average. It was fast. The study would get ordered, imaged, and within 16 minutes or so, you would get a result. That’s great; you could bring people through the ER [unclear: raw transcript reads “air”] quickly. Now, it’s not measured that way anymore. It’s measured in 45 minutes on average, an hour, or longer. Patients are sitting and backing up in the ED. It’s costing hospitals a huge amount of money, patients are getting frustrated, and delaying their care affects their health.

Something people might not know is happening: because it is so expensive to hire radiologists right now—there aren’t enough of them, so everyone is in a food fight for radiologists—salaries are going up while reimbursement for radiology has been going down for years. Expensive to hire, reimbursement coming down, massive mismatch—all of a sudden, a lot of radiology practices can’t survive anymore. They can’t manage their volume because they can’t hire people. They’re going to hospitals, and hospitals are paying subsidies to radiology practices just to keep them in business—millions of dollars. Capacity is a massive problem for patients, healthcare systems, radiologists, and radiology as a specialty.

CHRIS ST. JOHN 00:10:10 Yeah, for sure. I saw that you were speaking with Ian Weissman. We had him on the show a little while ago talking about this, particularly within the framework of staffing, though we didn’t get as much into the volume side of the story. It’s such an interesting moment, right? Because of course it’s horrible that people are backed up and that we need more eyes on these images. But as we have been thinking about this rise in volumes, I’m curious about the story behind it. A lot of this started happening around COVID, and while there is a downside, at the same time, theoretically there’s a lot of good coming from the fact that volumes are going up as more people get access to imaging. We’re overwhelmed in the trenches [unclear: raw transcript reads “dredges”], but at the same time, I hope the rise in volumes is indicative of a higher quality of care. Maybe, maybe not.

NINA KOTTLER 00:11:20 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 and we had new imaging that did new things. I actually think a lot of it is because patients are getting older. Patients in the U.S. are very sick; we’re not a super healthy population. With that, more people go through the system and more imaging happens. Imaging is very important for people who just never needed it before. Eighty-five percent of the time when a patient has a problem and goes to a clinician, they get an imaging study to help figure out what’s going on. So, 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. Some of it is because we need to get patients through the ED faster. We see a lot of studies being ordered not by the physician, but based on triage, or by nurse practitioners and advanced practice providers who haven’t had all the training or experience of other physicians and tend to order a bit more exams.

Is that the primary problem? I don’t think so. I think imaging is useful, the population is getting older, and more imaging studies are being done. We’re not going to solve this by cutting down the amount of imaging, even though that’s part of the problem, and we’re not going to solve it by increasing our workforce. The numbers tell a really important story: every year, the amount of imaging growth is an order of magnitude—10x bigger—than the amount of capacity being generated every single year.

In the past, we had cycles every four years: too few radiologists, then too many, going up and down. This is not a cycle. This is a direct line going up with a 10x difference over the next 10 years, as it’s expected to continue. We are projected to need another 15,000 radiologists when we only have 35,000 in the industry to begin with. It’s not possible to manage without technology.

CHRIS ST. JOHN 00:13:42 Let me dive into it. When we talked before we recorded this episode, you were talking about your early days using these clinical AI tools, where you had four different machines running like a sea of monitors, each running different AI tools with none of them connected. Can you paint a quick picture of what that version of your workflow and daily life was like, and then we can move into where things are now?

NINA KOTTLER 00:14:09 Where things are now is mostly still like that for most organizations. Who was it that said, “The future is here—it’s just not evenly distributed”? That’s exactly where we are, and it’s totally true for healthcare. In our organization, we’ve invested a lot in AI, so we’re a very early adopter. 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.

When you were giving that description, it reminded me of the movie The Matrix, where the guy is looking at 16 screens with green numbers. 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 (EMR), which has some information in radiology. We use our PACS, which shows the images and some data. We’ve got a RIS [unclear: transcribed as “risk”/”wrist”]. We’ve got all these different information systems that don’t connect to each other, or if they connect, it’s very superficial.

When I open up a study, I have four monitors, and they each have multiple different applications on them. Guess who has to consolidate all that information? A human. I look at the EMR for some things, I look in the RIS for something else, I take it all together, and then I concentrate on the images. But humans are terrible at integrating information; it’s just not what we’re built to do.

Computer systems are great at it. It’s just that we’ve never had a system that overlies everything and can translate all the information. Many of our systems were built in the 1980s and 1990s—EMRs a little later—which makes it very difficult. The standard of care today is a crazy number of systems with the clinician sitting in front of them integrating all that information.

Where are we moving to? We need all that information—it’s not that it isn’t important—but we want it presented in a way like Iron Man in his helmet: there are tons of data points it could give, but it gives the stuff he needs at that moment based on what he’s looking at and what he’s going to do. That’s the kind of system we need. What we’re moving to is away from individual AI tools that provide one detection output, to tools that help with workflow—integrating information together, summarizing it, and providing it so that you as the radiologist can continue looking only at the images, because that’s the patient and that’s what matters.

CHRIS ST. JOHN 00:16:48 It feels like a lot of the data you’re talking about—from the EMR and elsewhere—is relatively easy to find and acknowledge as having reached ground truth. Looking at things like dose or image quality, there are numbers where you can say, “Yes, this is right,” even when looking at summaries. But when looking at images, that’s where things get a bit murkier. What does the ideal workflow look like in your mind? Talk me through it a bit more.

NINA KOTTLER 00:17:37 Let me answer part of it at a time. It sounds like what you’re speaking about is that AI can hallucinate, and how do you know when the answer it’s giving you is right?

Sometimes it’s easy to verify because it’s a number that you can look somewhere else to verify. Other times, it’s making a decision about something, and how do you validate that? We know AI isn’t perfect. The best outcome occurs when you combine the human and the AI together so that both are better. If one works really well and the other works really well, but they don’t work well together, frankly you’re not helping the patient, because these AI systems are not autonomous. Autonomous means working fully by themselves without a human in the loop. Almost every single one of these systems is meant to be an adjunct, so you need the two to work better together.

Imaging is mostly an easy thing to verify, though not always. Let’s talk about the narrow AI solutions out there. Narrow AI solutions look at images, maybe even priors, and tell you if there’s something on that image that the AI sees—for instance, saying there’s blood in the brain. It’s kind of like what Google used to do when you brought up a picture of a cat and it said “cat,” or a picture of a dog and it said “dog.” Our imaging AI systems will say, “There is blood in the brain on this head CT,” or “I don’t see blood in the brain.” Another tool will say, “There is a hole in the lung,” or “There is not a hole in the lung.” “There is cancer,” or “There is not cancer.” They are binary yes/no decisions. It’s very different than what people probably realize and completely different from what you use day to day.

That’s state of the art right now for most people using AI. Those tools are additive on top of your PACS. 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 is. It’s helpful if it tells me a little bit more than just that there’s blood in the brain, because there are hundreds of images on a CT scan of the brain and it can be subtle. If it shows a picture or an image number of where it sees the blood, great—I can look and see whether it’s there or not.

What’s hard is quantitative imaging. There is some of this in AI right now, where it looks at the brain and says the volume of one part is X and another part is Y. We don’t manually measure volumes, so how do I validate if that volume is right? That’s when it gets more difficult.

Some things are very easy to validate; others are harder. We need to make sure the human and the AI system work well together. We have to be thoughtful about how to put them together so the human can validate the AI, because the human is ultimately responsible. Tools are not responsible for humans; humans are responsible for humans. You need the clinician to be armed with output from the AI, decisions it’s making, and transparency about how it works, so they can agree when the AI is right and disagree when it’s wrong—which is not always obvious.

CHRIS ST. JOHN 00:21:16 I’m curious about this yes/no binary approach to tools. Is this at Radiology Partners that you all are doing this, or is this everybody?

NINA KOTTLER 00:21:28 Not anymore—that’s everyone else outside of language. Anything with computer vision has to go through the FDA, and you cannot sell that tool until it is authorized by the FDA. The FDA has not authorized any of these generative AI tools that can do everything yet. I think that will happen next year, which means for computer vision, everyone is still using these narrow AI tools because they have been cleared by the FDA.

With those narrow tools, you could have 30 or 100 of them doing different things, though most people probably have 5 or 10 at most. So they’re only looking for 5 or 10 different things in your images.

Where we’re going is a place where the AI will look at everything in your head CT or brain MRI and dictate an entire report. That’s what we’re doing right now at Radiology Partners, and it is very different from what is happening elsewhere in the world. Because we are clinicians, we created this tool, and when you create the tool, you can use it under a medical exemption—if you’re a medical practitioner, you can try tools to make the system better, because you need clinicians to make the system better. We’re also rolling it out under a research protocol, which enables us to do something the industry is not yet doing. That’s why there’s a difference.

CHRIS ST. JOHN 00:23:00 By definition, would that be agentic AI, if we’re continuing to come back to buckets?

NINA KOTTLER 00:23:08 Let me talk about that, because I think everyone uses that term incorrectly as a hype term. If you say you’re using generative AI, multimodal AI, and agentic AI, maybe people will invest in your tool, so everyone calls it that. But let me define it better so it’s clear.

Remember how in the beginning you asked me what machine learning was? I said machine learning, or early versions of what we considered AI, was us programmatically telling the computer what to look for. In that scenario, we are making up all the rules. In healthcare, to do that well, you would have to predict every single scenario that occurs. There are too many edge cases to predict every action, so people started talking about agentic AI because it can be more predictive and adaptive.

An agentic AI system has to be able to do three things:

It has to take a high-level goal and break it down into rules that it creates and applies to reach that goal. For example, right now, to determine what study I read next, I have a system with built-in rules—if it’s a stat from the ER, read it first; if it’s an outpatient, read it later. It’s rules-based. If I change that to an agentic system, I wouldn’t give it rules; I would give it a goal: “Create a worklist that gives me the highest quality output and meets the best turnaround time so we get it done quickest.”

You have to give it access to different components of data and tools across your system. It needs to know what imaging studies are coming in, the schedule, which radiologists are working, the subspecialty of the study, and the quality of the study. It uses these tools to gather information and create its own rules.

It needs to be able to self-monitor. As it self-monitors, it realizes whether it is moving toward or away from its goal. If it’s moving away from the goal, it automatically adapts its own rules without requiring me to step in. It adapts the rules itself to move closer to the goal.

That’s what an agentic system is. If it doesn’t do those three things, I would not call it agentic. Agentic is not just a “better AI” using an LLM; agentic capability is something enabled within your system.

CHRIS ST. JOHN 00:26:31 Got it. I very much appreciate that. Self-monitoring is the piece of the puzzle I was definitely not cued into. Monitoring for AI drift seems like a crucial component across all sorts of different models, even outside radiology.

NINA KOTTLER 00:26:57 AI drift is a really good topic to talk about because the name itself is actually misleading. 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 changes.

You mentioned COVID before. Before COVID, we didn’t have disease that looked like COVID. If you had trained an AI model to look for pneumonia and then all of a sudden had COVID cases, it might not identify it well because it hadn’t been trained on it before.

In general, we say a model is either generalizable or not generalizable, and what makes it generalizable is the training data. If you train it on a small subset of data, it will do really well on that subset, but if you give it something different, it won’t necessarily do well. A generalizable model is trained on a much larger dataset, so it works across a broader patient population, exam population, scanner manufacturer, or protocol. All of those things are very important, and it’s another benefit of moving toward these much bigger foundation models, because they’re trained on so much more data than the narrow AI systems we’re using today.

CHRIS ST. JOHN 00:28:20 I get the scale of the data, but as you are trying to build these models, how do you monitor the quality of the data itself? Is the concept that a certain amount of scale will directly improve quality, or are you trying to balance scale and quality? What does that balance look like?

NINA KOTTLER 00:28:41 Scale does help with quality—the more data we give it, the better it tends to be, as long as it’s good data.

The difference between a narrow AI system and a foundation model is how you train it. Narrow AI systems are trained by giving them labeled data. Labeled data means a human [unclear: raw transcript reads “humid”] goes in, looks at the image—say, a chest X-ray—and if the narrow AI is looking for pneumonia, they literally circle the pixels with pneumonia or note that the image has pneumonia. That’s a human labeling the data. The data becomes structured, which is easier for a computer to understand, so you don’t need as much of it.

In contrast, foundation models are trained on a massive amount of data—millions of images. We can’t possibly label all of them; we don’t have enough radiologists to do the regular work, let alone the labeling work. You might think that without someone labeling it, it would be worse. But it turns out that if you give the AI the radiology report and the images, it’s pretty good at figuring out where the pneumonia is on its own.

The key is making sure the data doesn’t have really bad results. Not every radiologist is equally good with every system or study type, so you want to look at your data and exclude those, while elevating the radiologists who are really good. In theory, if you just created an AI system off a bunch of average reports, the AI would be as good as the average radiologist. But you want your AI system to be better than average. So you have to sift through your data to optimize it so model output is optimized.

We haven’t noticed a limit—as you give it more data, it continues to get better. It might be asymptotic, and there may be a point where it’s not worth the high cost of training when no one is paying for these systems, but in general, the more data you give it, the better.

CHRIS ST. JOHN 00:31:17 We’ve touched a lot on the data science, but I want to touch on the human element: trust in these systems, both human trust and patient trust. It’s a relatively high barrier. There’s societal distrust of AI for environmental reasons, bandwagon reasons, or artistic integrity—everybody has their own view, and they’re all valid. As a potential patient, I’m curious: how do you see the process of rolling these out to the public, and public perception in general?

NINA KOTTLER 00:32:09 A lot of people in the public are probably using these tools already, asking questions to AI because you get immediate answers, a lot of data, and it’s personalized. Some people share their data with public systems, while others are worried about that.

In general, people like having knowledge because knowledge enables action. If you want to take control over your health, having knowledge about your own health is important.

When we talk about deploying AI within a healthcare system—not a patient using it outside—there are many regulations and safety requirements we must manage through. Privacy is extremely important and has been a big public debate: how do you make sure patient health information (PHI) is safe and not used against them? We’ve had privacy rules for a long time in healthcare, so we’re pretty good there. The danger is if you share your information with an open system, then your information is out there. But if you’re doing it within something deployed by a hospital or FDA-authorized, we follow guidelines that have been in place for a long time, so I’m less concerned about patient privacy within healthcare.

AI has immense potential value in healthcare in particular because it’s one of the only things that can solve this capacity problem and enhance quality. I think it will help transition us from population health metrics—treating a broad population—to precision medicine: deciding what’s wrong for that specific patient based on their genome, genetics, and labs combined.

We’re also moving from diagnosing what already exists to predicting what’s going to come, which is much more impactful. Those are the kinds of things AI can do. Within our current system, we have enough guidelines protecting patients that I see a massive amount of benefit in moving forward. I hope patients understand that what happens outside of healthcare has a different set of rules than what happens within healthcare.

CHRIS ST. JOHN 00:35:09 What you said about people craving knowledge really resonated in this context, especially when talking about predictive findings. There’s something human and tempting about being able to know what could or will happen. You saw it with genetic cheek swab companies like 23andMe and all those DNA tests—people were so curious to learn about themselves that they sent off their DNA without reading the fine print, just happy to know if they were predisposed to a condition. It’s a really interesting point. But what about a skeptical radiologist who is being told they need to change their workflow?

NINA KOTTLER 00:36:12 That’s hard because change is scary. No human likes change, especially in an environment that is extremely stressful. Right now, we’re completely overwhelmed with volume, and there’s a lot of legal liability—we are responsible for patients’ lives. When you start to make a change, it adds uncertainty and extra work.

I’ll give you an analogy. Humans cannot do two things at once requiring frontal lobe executive functioning. But why can we ride a bike and think about what we’re going to say on a phone call, or drive and do something else? Because that second function is built into our basal ganglia, the part of our brain that enables us to perform functions without using our frontal lobe.

The way a radiologist functions in this super stressful environment—all those clicks, integrating information from one place to another, scrolling through an exam, knowing where to look—has been ingrained over years. I’ve been in practice 20 years, and you get really good at it; it’s like riding a bike. If all of a sudden you turn the handlebars so that turning right actually goes left, you suddenly have to think a lot about what buttons to push, which makes it harder to do other things. That’s scary because ultimately we don’t want to harm patient care, so any change we institute is hard.

Does that mean we shouldn’t do it? No, absolutely not. We need to continue changing and evolving, but in a safe way. Everyone has a different level of risk aversion: some are excited about new technology and want to adopt it right away, while others are not. It’s about managing change across that spectrum—making sure the people who are excited aren’t using it inappropriately, and helping the people who are scared start using it in the right way because it will make them better. That’s the whole point.

Education is probably the most important thing, along with transparency. How do you make sure the AI provides the information the radiologist needs to understand how it’s making decisions? A lot of that isn’t happening from vendors today. We have to tell vendors we need information that makes the human and the radiologist better together, and that should be part of the regulations and requirements for these tools.

CHRIS ST. JOHN 00:38:56 Different types of healthcare facilities are going to have different levels of access to these tools. We see it with scanners and staffing—smaller rural facilities versus big academic centers. What would you say to smaller facilities that are doing a decent amount of volume, are underwater, and don’t have access to these tools? What advice or direction would you give them in this interim period as things ramp up, especially as we try to safeguard against burnout among our current radiology professionals?

NINA KOTTLER 00:39:47 We need to not be afraid of this; we need to jump in and use it now. Hospital systems, practices, and groups that don’t are going to fall behind, while groups that use AI will outperform them. They might think in the short term, “I can wait,” but that’s going to hurt them in the long term.

Jump in and get involved. The people who get involved early are the ones who get to make decisions about where these things go. It’s really important for clinicians and healthcare systems to make those decisions, not just vendors.

The hard part is there’s no return on investment from a payer standpoint—CMS, Medicare, and private payers are not paying for these tools. How do smaller hospitals afford them? Each AI tool has to have its own ROI, and that ROI determines who should pay for it. If the ROI is that the radiologist becomes much more efficient, then the radiology practice should pay for it. If the ROI is that more patients go through the system, get better downstream care, and the hospital saves money because patients get through faster with a better experience, then the hospital should pay for it. If an AI tool doesn’t have an ROI, it’s dead in the water.

There’s something we need to be doing called AI governance. AI governance means overseeing the AI to make sure it’s safe over time and ensuring the end user uses it correctly—clinical governance. Not everyone has the dollars to afford that. I’ve been speaking to people in government and societies, and I think we should start talking to CMS about not reimbursing the AI itself, but instead reimbursing the clinical governance overseeing these tools to make sure they’re deployed appropriately and monitored over time. That would be my request.

CHRIS ST. JOHN 00:42:08 How do you think about quantifying ROI in this context? Obviously, in some cases, you could say, “We’ve done this many more exams since implementation—that’s ROI.” But for so many clinical tools, ROI feels soft and nebulous. How do you think about that?

NINA KOTTLER 00:42:35 You have to define exactly what the ROI is because the C-level executives—the CFO and CEO—are buying these tools, and they care about the exact dollars. It’s not, “There’s an ROI because your patients are going to love you.” It’s, “You’ll get this many more people through the emergency department and have this many more admissions.” I kind of hate talking about this because it sounds like a business, but hospitals have to stay in business, even though as clinicians we care about patient care.

CHRIS ST. JOHN 00:43:18 Capitalist society—we get it. We exist in it; it’s part of the conversation.

NINA KOTTLER 00:43:24 It is. Most hospitals are paid through what is called fee-for-service: do more services, get more fees. If they do more procedures—and we want them to do the right procedures—or bring patients from one part of the system into another, there’s ROI.

For example, we have something called breast arterial calcification, where you can identify calcifications in the vessels of the breast when doing a mammogram. It turns out that has predictive value for a patient’s risk of coronary artery disease. If you take patients being screened for breast cancer and refer them to cardiologists—since many women present with their first heart attack with no warning signs—you’re bringing patients into another part of the system, which adds revenue. You have to break it down and think about how dollars can recoup costs so we can put these great tools into place.

CHRIS ST. JOHN 00:44:26 As much as I hate to end with capitalism, that’s about all the time we have today, although I could keep talking to you for hours. Dr. Kottler, thank you so much for coming on the show and talking with me today. It’s been amazing having you.

NINA KOTTLER 00:44:43 You’re very welcome. I appreciate it, Chris. It was a very fun time.

Questions from This Episode

What is the difference between narrow AI and foundation models in radiology?
Narrow AI consists of targeted algorithms cleared by the FDA to perform specific detection tasks, such as identifying intracranial hemorrhage or pneumothorax. Foundation models use multimodal architectures that process language and vision data across large collections of studies. These larger models can evaluate complete examinations, synthesize clinical context, and draft diagnostic reports.
What makes an AI system agentic in a healthcare environment?
An agentic AI system must take a broad operational goal and break it into self-generated rules, access data tools across connected systems such as PACS and EMRs, and continuously monitor its own performance to adjust those rules. This allows the system to manage variable clinical workflows such as worklist prioritization without manual reprogramming.
How do healthcare organizations establish a return on investment for radiology AI?
Return on investment depends on the tool’s clinical or operational function. Efficiency tools that increase reader speed sit within practice budgets, while tools that improve emergency department throughput or identify secondary conditions can create value for hospital systems. Organizations must align technology costs with the group receiving the financial benefit.