Rethink Imaging

EP 42 • April 16, 2026

The Cognitive Ceiling: Why AI Must Support Radiologists, Not Replace Them

Featured Guest

Julie Bauml, MD

VP of Imaging and Clinical Informatics • 4DMedical

Listen Now46 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

The field of radiology is at a breaking point, facing a massive “volume crush” that threatens to push professionals toward early retirement. Dr. Julie Bauml, a board-certified radiologist who transitioned into clinical informatics, joins the show to discuss why technology must be reframed as a tool for augmentation rather than a replacement for human expertise.

The episode explores the “cognitive ceiling,” a biological limit on how many high-level decisions and visual inputs a brain can process in a day. Dr. Bauml shares sobering anecdotes of AI tools implemented without radiologist input, resulting in poor workflow integration and even decreased reimbursement. From the “Minority Report” dream of hands-free, spatial computing to the nuances of “ground truth” in clinical data, this conversation is a tactical roadmap for keeping radiologists meaningfully in the loop while automating the rote tasks that lead to burnout.

Host

Chris St. John

Host, Rethink Imaging • Imalogix

Featured Guest

Julie Bauml, MD

VP of Imaging and Clinical Informatics • 4DMedical

Watch the Episode

Key Takeaways

AI should focus on automating the non-doctoring tasks (chart searching, window alignment, lesion measurement), not replacing diagnostic interpretation.
There is a biological limit on how many high-level decisions and images a radiologist can process in a day. Efficiency tools that only increase throughput accelerate burnout rather than relieve it.
AI tools deployed without radiologist input consistently create downstream problems: poor workflow integration, unresolvable false positives, and in some cases reduced reimbursement.
Data quality and diversity matter more than data quantity. Training tools on larger but noisier datasets does not produce better clinical performance.

Full Transcript

[00:00:05] Chris St. John: Burnout and pressure in radiology, I feel like, are coming up a lot these days. You can’t just use these tools to perpetually increase volume at the same rate as the burnout.

[00:00:16] Julie Bauml: There is a cognitive ceiling on how much a person can read a day, and we’re nearly almost there.

[00:00:22] 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. Alright. Welcome back to Rethink Imaging. I’m your host, Chris St. John. Today, I am joined by Dr. Julie Bauml, a board certified radiologist who made a rare and intentional move into clinical informatics and AI. Julie trained in radiology and MRI, practiced in both academic and private settings, and then went back to do a clinical informatics fellowship at Vanderbilt, and she is currently still reading cases for Imogen. Julie is currently the Associate CMIO at Hoppr, adjunct faculty at Vanderbilt, and her work focuses on AI as augmentation, not as replacement, with a big emphasis on data quality and keeping radiologists meaningfully in the loop. Julie, welcome to the show.

[00:01:23] Julie Bauml: Thank you, Chris. It’s good to be here.

[00:01:25] Chris St. John: It’s super exciting to have you here. I feel like I talk with a lot of people about AI and imaging, and I just felt your whole view of it was honestly quite refreshing and exciting, and maybe a little bit skeptical even to some degree. Is that fair to say? Maybe that’s not fair to say.

[00:01:46] Julie Bauml: No. Sure. It’s fair to say. It’s more, you know, cautious optimism. Right? Because we’ve all seen a lot of promises come and go for advancements in AI and radiology. And so now we have learned to test the waters a little bit before we go all in.

[00:02:01] Chris St. John: Yeah. So before we get too deep into the meat and potatoes of what you do, I’d love to just hear a little bit more about your career transition. You had a very successful career in clinical radiology, and yet you chose to take the step into informatics and AI. I’m curious, what kind of drove that transition for you?

[00:02:24] Julie Bauml: Yeah. So I think, ultimately, it boiled down to just sort of frustration with the system as it was. I had worked in academics. I’d worked in a few private practice academics, full private practice, kind of the whole gamut. And all of those reading environments, working environments had pros and cons to them, but the steady through line of the volume crush really starting to come close in on people was everywhere you went. And I think I maybe just saw a little bit ahead of the curve on that to try to come up with a way that I could practice again without losing my spark, without losing what made me love radiology, and feel like I could retire early. I could complain, I could be like, oh my god, the list. Or I could try to do something. So it was a pretty strange thing to do, being a full attending for years and making a full salary and then saying, no, I’m gonna go back and do two more years of fellowship. So I did that, and I went and did it at Vanderbilt because I thought that was the best opportunity for me. The Vanderbilt program is really strong, and I’ve been really happy with that choice. It’s actually given me a way to continue practicing and try to be part of what I think is a really necessary solution for the issues of radiology today.

[00:03:58] Chris St. John: Okay. Beautiful. Do you mind touching a little bit on what this solution is that you’re kind of talking about?

[00:04:05] Julie Bauml: Well, I wish I knew entirely, because it’s a very complex and interdependent ecosystem, medical imaging is, and of course all of medicine is. And you’ve seen large non-medical company after company try to dip their feet in the water and realize it’s not that easy, and then have to pull out real fast. I’m thinking of Walmart and others that had to really scale down their attempt at coming into clinical medicine, and they realize how complicated it is. But the ultimate solution is probably going to be multifaceted, because all of the problems are. The volume is really, really high, but it’s also a fee-for-service structure. So you could decrease the volume, and everyone’s dollars will go down, and people won’t be getting the imaging they need. So okay, that doesn’t fully work. You could try to make us more efficient, but there is a cognitive ceiling on how much a person can read a day, and we’re nearly almost there in some ways. And there are other avenues where there might be a little bit of give with some AI tools, with some non-AI tools, all of that. There’s reducing truly unnecessary scans. It’s this whole arrows coming from every direction sort of thing. Whatever we do, we need to be thoughtful about where we gain efficiency, not just saying, okay, now read more studies, because I guarantee the radiologist won’t feel less burned out at the end.

[00:05:33] Chris St. John: Right. You know, it’s funny, I’ve never heard anyone phrase it as a cognitive ceiling. I really like that phrasing. Because when we frame it around capacity or even within the context of burnout, it feels like a more nebulous barrier. But when you phrase it as a cognitive ceiling, it really does, to me, frame it as something that does have a much harder limit than I think we often talk about it.

[00:06:01] Julie Bauml: Yes. And I actually think it’s beyond just cognitive, which would be for all of medicine. It’s also for radiology, like a visual. How many images can you take into your cortex in rapid succession throughout the day? There’s a limit to that. It’s not like if we could be infinitely efficient, we could just keep our eyes open. There is a limit to that as well. Maybe it’s a little higher than it is in some of the other subspecialties of medicine. But even in non-radiology, I’ve seen things happen where administrators say, oh, we have this ambient AI software, and it’ll make you 30% faster. It’ll write your notes for you while you’re in there. You’ll have better patient interaction. It’ll order your stuff, and we’ll give it to you. But to pay for it, you need to see 30% more patients per day. And so every single person will say no. Because there’s also a cognitive load. Well, you’re not decreasing my burnout at all. You’re just increasing my efficiency. That’s not to say we shouldn’t push on such tools and gain efficiencies where we can. But if we don’t give at least a little bit of that back to the end user, they’re not gonna thank us.

[00:07:16] Chris St. John: Right. You can’t just use these tools to perpetually increase volume at the same rate as the burnout. Burnout and pressure in radiology, I feel like, are coming up a lot these days, understandably. So I’m curious, what do you think folks outside of the field still misunderstand about a contemporary radiologist’s day to day?

[00:07:44] Julie Bauml: Yeah. I think one of the big things that people don’t understand about what the radiologist is doing is that they’ll come into the reading room and they’ll see you sitting at a computer, and they’ll be like, oh, great, you’re not doing anything. And then they’ll just spiral off all these questions. You’re literally in the middle of deciding if some kid has appendicitis and trying to communicate it to the ER. As they see you in the chair, they think you’re relaxing all day, where in reality, you’re trying to stare into the void of images and make thousands of high-level cognitive decisions a day that directly affect people’s lives. And even though it looks like, oh well, they’re just sitting there, you’re really feeling a lot of that intense pressure. We’re not just ever during the day, really, with how volumes are, just casually chatting or relaxing. We’re just go, go, go, moving as fast as we can. It’s not necessarily as leisurely as our fancy chairs with the ergonomic backs would make it seem.

[00:08:42] Chris St. John: Yeah. There’s parts of a radiologist workflow that are cognitively draining, emotionally draining, and then also just time consuming. Distinguishing between what’s what when trying to address the volumes we are starting to see in imaging, I think they’re important distinctions to be making.

[00:09:07] Julie Bauml: Oh, for sure. It’s so important to distinguish between those things. And you’ll see a lot of the earlier tools that came along, through no fault of their own, because AI and ML starts progressing and people start to make things that they can make. And the use case is maybe secondary because it’s just a matter of what can feasibly be done at that point. I’m talking ten years ago, and you’ll have all these point solutions come in, and the focus was so heavy on diagnostic tasks for the user.

[00:09:43] Chris St. John: Right.

[00:09:43] Julie Bauml: Is there or is there not a pneumothorax, or is there not a head bleed, or what have you? And I think the generous interpretation of that is that’s all we could do at the time, and a hammer looks for a nail. And the slightly less generous one is people were pretty darn convinced they were gonna replace radiologists, so why care about the workflow? But you’ve seen a big shift away from that kind of mentality in more recent years as people realize replacement is very far down the road. What you were saying about the different types of tasks and thinking about reducing the volume crushing, the overload on the radiologist’s mental capacity, is really important. Because there’s so much to be had with things that don’t require me generating a workable report, as long as I get the data in there in a very terse way and we have a system that can generate a report or generate my impression for me. That’s not the top of my license. That’s not the doctoring part. Searching in the chart for the history to try to figure out why they ordered this exam, getting all the windows aligned, that’s not it either. Measuring little lesions is not it either. These are the tasks that we need to focus on if we want to reduce that cognitive burden and allow radiologists to move a little bit more quickly, because we do have to pay for it. We do have to make things more efficient to pay for any tools we’re generating. There has to be a balance in there where they’re able to go a little bit faster because they’re feeling good, and they’re feeling less strain from each case they’re going through and focusing on the rote things. The diagnosis part, that’s my job. I’m happy to have you check me on the back end or point out things that I might have missed, but the meat of my job is the diagnosis. So anything we can do that increases eyeballs on the right screen, I think, is really important. And what’s the right screen? The report generation screen is not the right screen. The history screen is not the right screen. It’s the images of the patient where I’m actually looking and deciding what’s wrong with them.

[00:12:21] Chris St. John: Yeah. Let’s say you have an unending pile of exams to read. Totally theoretical. Never real. You have this massive nonstop pile of exams to go through. Let’s say you have some efficiency tools in place helping with your dictation, helping with the history, all of that. How do you see your own relationship to your own cognitive ceiling as it’s tied to burnout but trying to go through all, are you able to draw that line yourself at times, or is that not even possible?

[00:12:55] Julie Bauml: So depending on what sort of practice environment you’re in, you may or may not feel like you have the opportunity to draw that line. If you’re the only radiologist reading stat or emergent cases for a group and no one else is on call, I have had this happen to me so many times over my career. You gotta keep going. And you can see on the list that you’re dealing with a complex cervical spine, but guess what? There’s four abdomens that just popped up right after it. You have a turnaround time of forty-five minutes to an hour from when any study hits your list. You don’t have an option to spend an extra five minutes. You gotta go. You gotta push. Whereas if you’re in my reading environment, for Imogen it’s I get paid for what I read, and it’s a very large practice with a lot of lists. You can really only do this in a big group or possibly a tele-group where you have enough people that nothing’s gonna go unread and people can go at their own pace. I prefer that so I can spend the requisite time per patient that I feel is appropriate without hurting someone else. Either the patient’s not getting it or, if we’re in a practice, we’re all reading off a common list, and if I’m going slower, everyone else has to go faster. But the truth is, there’s a pay drop-off for that. This is the ecosystem I’m talking about.

[00:14:14] Chris St. John: You have been known to say radiologists need a reasonable seat at the table. I feel like that is something I’ve seen come up a lot surrounding you and what you’re talking about. I’m kind of curious, when it comes to technology, are you seeing radiologists getting their reasonable seat at the table, or not really yet, where it feels most relevant?

[00:14:39] Julie Bauml: I think they’re getting more of a seat at the table. I work for Hoppr. We have six or seven radiologists for our company, which is amazing. And I think when you look at a company and you’re roaming the halls of RSNA or whatever, and you go up to their booth and they don’t have a single MD in sight, that tells you something about their philosophy. Are they listening? Are they laser focused on replacement? These are signs sometimes that they are. Not a hundred percent of the time, of course. Radiologists are getting more of a seat at the table because a lot of the earlier ventures that did not have them there did not work. So now they’re like, oh, it turns out this medicine thing is really complicated. It’s not just an ML solution where we implement it. Even if someone came up with an extremely accurate AI tomorrow that could do a thousand diagnoses, it doesn’t exist. But if they did, it wouldn’t replace anyone tomorrow. Everything needs to be adapted and adopted into the workflow and clinically proven and taken in and adapted for all of the nuance for the particular practice environments. Things move slowly in this space. The early talk alienated a lot of radiologists. Some companies would say, oh, we want more radiologists to be involved. Radiologists are very heavily involved on the AI side, things like the ACR and professional organizations, but maybe not as much in industry as they could be. They’d say, oh well, we want the radiologists, but they’re not interested. A lot of that bad taste in their mouth was from, well, you guys were claiming to replace us and that we should stop training people since 2014.

[00:16:31] Chris St. John: Right.

[00:16:31] Julie Bauml: I didn’t exactly want to put a hand in my own demise. Although frankly, I don’t have a dog in that fight. I’d be happy to be replaced by a perfect AI and not feel guilty about enjoying my relaxation. But most people don’t feel that way. And of course they don’t, because you have people coming out saying, you know, the ones everyone knows about, the Khoslas and Hintons coming out and saying, oh, it’s tragic they’re still training radiologists. This was ten-ish years ago at this point. Oh my gosh, they’re not gonna have jobs. Five years, they’re gonna be obsolete. Five years. That was the quote. Five years. Do you know how long radiology residency is? Five years.

[00:17:14] Chris St. John: For the audio listeners, we were both just holding up five fingers.

[00:17:17] Julie Bauml: Five years. So what are you telling people about our history? But I think a lot of that bidirectional, you guys think you can replace us, and them saying, you know, human in the loop is poop or whatever the spicy bro take of the moment is. A lot of that has died down as each group sees that it needs each other more. And so radiologists are able, if they want, to get a larger seat at the table. People are listening more than ever before, because the hiring crush, the volume crush, it’s so severe, and radiologists are tapping out of the system at higher rates than ever before. Early retirement, going from full time to part time to kind of deal with, is there a way to deal with your cognitive ceiling? There’s no great way. These are the things that people are choosing just to preserve their sanity. So for the first time ever, us not being happy matters a little bit, so people are listening a little bit. But it’s not really about us. It’s only about us so far as the radiologist being happy and able to practice longer and more thoroughly and better is good for the patients, because that’s really what matters.

[00:18:35] Chris St. John: Well, yeah, of course it’s about y’all, because it’s an ecosystem. You can’t have, all of a sudden, there’s no more wolves, and the deer population is, you need every player in the puzzle to keep the ecosystem balanced and moving. So if radiologists are burning out at an insane rate, the system just can’t continue to function.

[00:19:00] Julie Bauml: To the point of when you don’t include them, what kind of stuff happens. I have a colleague whose group was dealing with a horrible chest CT backlog.

[00:19:11] Chris St. John: Mhmm.

[00:19:11] Julie Bauml: Administration is like, oh, these chest CTs are going weeks without being read, and we’ve got this huge backlog of chest CTs. So they purchased AI pulmonary nodule detection tool software. They assumed that must be the bottleneck, and giving you this tool, the company claims it’ll increase x efficiency or what have you, and there you go. They didn’t involve the radiologists in the decision, so they also didn’t bring the tool in and integrate it in a way that worked with the existing workflow. It was poorly integrated. The outputs couldn’t be easily edited or overwritten. So even if they disagreed, the findings stayed there. And if the person doesn’t read the report and just looks at the PDF or the input from the company, they’re gonna get confused. It just caused all this confusion with the downstream providers, or recalling nodules because they didn’t consult with the radiologists to set the sensitivity point, which is something they could have done. And so now people want to read the chest CTs less than ever, because now you have to go through this horrible tool. And on top of that, they go, okay, we’re actually gonna decrease your reimbursement for the chest CTs by 50% because you should be 50% more efficient because we bought you this tool. You are welcome.

[00:20:33] Chris St. John: Oh my lord.

[00:20:34] Julie Bauml: If you don’t involve the people who are the end users, this is always a tension. It depends on the practice setting. If it’s a private practice, they’re purchasing the tool for themselves and using it themselves, it’s one thing. If it’s a practice where the equipment or the PACS system is owned by an administration from a different system or a hospital system, this kind of thing can happen if they don’t talk to each other. And that’s on such a small level. And then when we build large tools and we don’t talk to the radiologist, we all magnify those mistakes.

[00:21:04] Chris St. John: I’m curious. We’ve touched on augmentation versus replacement several times over the last few minutes. You’ve always been very clear quite publicly that AI should be about augmentation and not about replacement. I’m curious, to you, what could augmentation look like successfully?

[00:21:24] Julie Bauml: Dream scenario, magic wand scenario?

[00:21:28] Chris St. John: Yes, exactly. That’s what I’m asking for.

[00:21:30] Julie Bauml: Yeah. In my dream future state, it would look kind of similar to Minority Report, which is a movie from I don’t even know how many years, twenty years ago. Ten?

[00:21:42] Chris St. John: I think Minority Report is like 2005 or something, if I had to guess.

[00:21:45] Julie Bauml: Twenty years ago. Where they already had envisioned this, they’re just swiping, there’s nothing in their hands. I’m not tip-tapping into the computer. It’s tracking my visual, where I want to look. I’m scrolling almost with a nod of the head. I can just ask for the information and it comes right to me. And all of Tom Cruise’s attention, when he was looking for whatever he was looking for, it’s been a while since I’ve seen it, is on the images at play, and he’s looking for very subtle things in the video feeds. That’s what it should be for us. All of our attention should be on the visual. There should be nothing about clutching a microphone and clutching a mouse, and trying to, oh, shoot, I was just dictating, but the dictation box wasn’t active that whole time, and I just lost all of that information. I could see it much more like, okay, scroll through it for me. Okay, stop. I see a lesion. It’s a whatever I think it is, because that’s the part that is my job. Measure it and put it into the report for me. That’s the part you can offload. And then at the end, hey, anything you saw that didn’t make it into my report, show me. Yes, that one’s true. It’s not important, but you can go ahead and print the body of the report. That one, no. Next case. That’s where I could see true augmentation. Some things are rare. I don’t care how much data you have to train something, it’s not going to be able to replace a human who can read a textbook and see some rare entity one time and then recognize, this is that thing, I gotta look it up and get this right. Maybe someday you give it all of human knowledge and none of us have jobs in any profession anymore. But I think that would be pretty far down the line from where we are now. So where does that leave us? It leaves us with augmentation. And in the preceding example we’re talking about, where we gave you this tool that made it worse and then told you were gonna pay you less, that kind of stuff happens all the time, but they were violating the central tenet of informatics. I went back and did a two-year informatics fellowship, not imaging informatics, clinical informatics. And the basic tenet of that is supposed to be: human brain plus computer better than human brain alone. But kind of inherent to that is sort of a reflection of do no harm, which is, at the very least, don’t make it worse.

[00:24:21] Chris St. John: You’re welcome.

[00:24:22] Julie Bauml: At the very least, human brain plus computer equal to human brain. Please please don’t make it worse.

[00:24:30] Chris St. John: Yeah. We’re gonna put a pin in this Minority Report scene, which also lives iconically in my brain. Seriously, it was a very important, I think, portrayal of the progression of technology for a lot of people. But I want to touch on ground truth a little bit, specifically within the context of medical imaging, because I think defining it is way harder than people assume. Would you mind defining ground truth a little bit and talking about why it’s so hard?

[00:25:03] Julie Bauml: Sure. It’s really hard. When you’re a machine learning engineer and you have a labeled dataset, a test set of values, it’s binary for the most part. This label, true. This label, false. The real world is unfortunately way more nuanced than that, because who decided true or false? And ground truth, it sounds like something immovable, but in the real world, it’s something that we negotiate, that has multiple entry points. So when we talk about developing AI tools for medical imaging, often the ground truth is the radiologist. The radiologist is human, so we don’t always even agree with each other. We don’t always agree with ourselves if you show us the same patient. Six weeks later, we’re not gonna give you the exact wording of the exact same report, and we don’t use language in the same way. So it’s extremely complex, even if you’re just dealing on the radiology level. Because of what’s available in terms of large amounts of data, it’s mostly image-report pairs. This is a CT exam, this is the report that was generated, and that is the label that you have. And that is, in essence, a weak label. It’s full of uncertainty, hedging, institutional variation, human error. And so that’s what we have. That’s what we go with for the most part. I don’t know how we can expect tools to do as well as us when the ceiling is an average of us. And then if you want them to do better than us, we need to have better information than we have. I liken radiologists, especially in the current system where we have very little time to follow up any of our reports and find out what happened to that patient, similar to a quarterback who throws the football but never looks to see if someone catches it. He’s crushing it. Oh, this person has whatever tumor, good luck, I hope someone catches that. And in essence, that is how we mostly practice. Yes, we have things like peer review and other things, but we’re not really getting any kind of full rundown on here’s all the cases you were right about, here’s the ones that you were slightly wrong about, and who would determine if we’re right or wrong anyway. And that gets into some of the so-called higher-level gold standards for a diagnosis or a process that we might call pathology reports. Okay, fine, pathology report is the gold standard. Alright, what about in mammography where you have an extremely suspicious lesion? Pathology says it’s benign. The whole idea of concordance versus discordance comes in with that. Why? Because we know that if it looks super suspicious, no matter what that biopsy comes back, that person’s going to surgery and having it removed anyway. So is pathology even a gold standard? Not necessarily. Patient outcomes, how long did the patient live? That’s a kind of signal. Did they have other lab tests that suggested your diagnosis? Maybe said that they have rheumatoid arthritis, the lab tests also bear it out, and they responded to treatment for rheumatoid arthritis. These are strong signals. We’re getting multimodal signals then for a ground truth. There’s layers of ground truth, and there’s different levels of signal from each of these layers that we have to take in. And it’s context dependent too.

[00:28:49] Chris St. John: Well yeah, and it makes me think about the variability between radiologists, between users, their own truths, and what level of variability is acceptable before it starts to mess with these AI models anyway.

[00:29:10] Julie Bauml: What do you mean by that? In terms of level of variability in the reports, or level of, as the training data?

[00:29:18] Chris St. John: Yeah, I guess that’s what I’m going for.

[00:29:20] Julie Bauml: Maybe that doesn’t make sense. For sure it is. Everyone has, you know, people think that radiology is just like a lab test sometimes, where they go, I shoved radiology in and diagnosis comes out the other side. Yes, no. That’s not really how it works. Radiologists operate like most tests do, on an ROC curve of sensitivity versus false positives. And if you called every single mammogram negative, you would have a very high accuracy, actually, because only a few in a thousand will have cancer, I think three or four, maybe that’s too high, it’s been a while since I’ve been doing them. But anyway, you’d be like, oh my gosh, look at my accuracy. So it’s context dependent in that regard too. And then the wording that we use can be extremely variable. Even if you’re right, but you’re describing it in a way that can’t really be filtered out and be understood by any text extraction or whatever you’re doing to feed into a model, that’s very difficult. And I’ve actually been thinking about that a lot, the wording, the standardization in recent days, because I was gonna say, radiology is more like an art in that you take all this stuff into account and give your best guess at the diagnosis. And it’s more like an art than a lab test, but it really is somewhere in between. But I’d be a hypocrite to say that because I want to turn right around and say that we all need to start fighting for standardization so much.

[00:31:10] Chris St. John: There’s so many different pieces at play. I think at the core of what you’re saying is, language, the way we use it, the way we speak it, shapes and defines our experience of the world. Cognition is very often based in language for us, and the language that we speak and the words that we think, those definitions explain how we see the world. And so then you move it into this medical imaging context, you have the same sort of framework of language defining what you’re seeing on the screen. But because of everybody’s variation in different relationship to language, you’re gonna get different outputs to some degree, and there’s nuance to that.

[00:32:01] Julie Bauml: There is a lot. And that’s why, you know, if the smartest people in the world trying to extract out the diagnosis and label all the synonyms to the best of their ability are struggling, that is, well, whatever, my report isn’t for them. It isn’t. The reports are being generated to take clinical care of the patient. But if people who are very financially and mentally motivated to understand you can’t understand you because of the level of variability to make an AI product, well, maybe the patients and the referring clinicians can’t understand you either at that level of variability. So I used to be like a dyed-in-the-wool, you can force a template on me in my cold dead hands sort of, I’m gonna free-dictate everything I want to. But the more I’ve seen of what that level of variation leads to, not just for developing tools, but for patients and understanding, I’ve warmed up to the idea of, okay, I will try to learn a more standardized system. I’m a radiologist, so I’m biased. We’re just throwing hands, being like, it’s this fault, that fault, the volume, the this, the admins, whatever. I think this is part of something we can take on and say, okay, it’s going to be better for the future, for our patients, for everyone, if we can decide on more standardized language in general so that everyone can understand each other. And I don’t want to turn into that meme where they’re like, oh, and thus the seventeenth standard was formed of terminology or ontology to encompass them all. But just in general, we were able to do it for mammography, and that has been very good for the interoperability of mammography and taking images and reports from one place to another place and still being understood by the mammographers there. Granted, that’s a small area of the body with almost like a binary system of what you’re looking for, but we can do better across the boards. This is something I’ve changed my mind on a lot recently.

[00:34:07] Chris St. John: I’m curious how. I’m making a bit of a jump here, so maybe it’s just switching gears. I’m curious where you stand about the debate between data quantity and data quality. If you’ve heard people out in the world saying, oh, we just need more data, we just need more data, more data, more data, I’m curious if anything worries you about that approach.

[00:34:38] Julie Bauml: Yeah, it doesn’t work.

[00:34:41] Chris St. John: Okay, say more.

[00:34:44] Julie Bauml: Well, like we were just talking about the complexity of ground truth within just deciding what it is, the quality of data is important in that way too. You’re gonna have this variability, not just in image quality, but the report quality. Are you hitting all of the major groups being represented, age, gender, rural divide, city, diagnoses being represented? And I guess now I’m maybe falling more into my thoughts of my work at Hoppr making foundation models, making VLMs, making things that understand a wide swath of imaging rather than just a pinpoint of it with one task to do. Since you really need to think about getting multimodal data, getting not just the image pairs, but also maybe supplementing that with human-derived labels for what you’re specifically trying to do. Get the pathology reports if you can, get the EHR data if you can, to really supercharge it. So I guess that’s more, not just data quality, but also diversity of data. But data quality is really, really important to making sure that, at the very least, what we do have is vetted for it being what it says it is when we figure it into the model, and the tags being what they say they are, and the images being processable. It sounds simple, but it’s actually not. When you’re just buying these large gobs of data or getting them from open sources, they’ve sort of just thrown it together. It’s just easier to just be like, yeah, here’s everything in our PACS system. But to go through and actually catalog that in a way that is digestible and useful, I think, is really important. And yes, people do have a tendency to say, just more data. Oh, it’s not good enough? Well, we tried it on a million. Let’s try it on two million. Let’s try it on three million, ten million, let’s keep going.

[00:36:39] Chris St. John: Right.

[00:36:39] Julie Bauml: And the CTO of our company liked to talk about that and make it a metaphor to just trying to bake a cake and responding to every problem by being like, more flour.

[00:36:53] Chris St. John: Right.

[00:36:54] Julie Bauml: More flour. Just more. Just, like, buy a whole shelf full of it. This cake isn’t working. And it’s like, bro, you need eggs and butter. You need some other stuff in there, and you need to make sure the quality of your ingredients is right. Or just adding more ingredients that are not being treated in a way where you have data integrity and you have traceability, and all of that, you’re not gonna get too far, I think.

[00:37:24] Chris St. John: Right. Okay. Would you mind just talking me through a little bit then, task-specific versus foundational models, and kind of your thinking and your ethos around them?

[00:37:34] Julie Bauml: Sure. So task-specific are usually built off of a CNN or something like that, and they’re truly trying to do one thing to the best of their ability. Often, it’s something acute because to make any use out of it, you would want to pick up something that you could do, like, a workflow orchestration with or something like that, to make it worthwhile and be like, oh hey, read this patient next, they have a pneumothorax or they have a head bleed or whatever it is. And that’s nice. It doesn’t save a lot of time. It might have some patient safety implications where we read those studies quicker, depending on what the actual workflow is of that given practice and whether or not there’s any bandwidth to move things up and down the list. But it’s a nice check on us. But when we read, we have to take all of the visual data into account and interpret in the clinical context. So this requires a much more thorough understanding, not just of what is a pneumothorax, but what is a CT? What is an X-ray? What is the type of medical imaging data that we’re looking at? What are the different things that go into that, not just in terms of the diagnosis, but also the metadata and everything like that? What kind of a study was it? How much radiation? What was it ordered for? All of this stuff, trying to understand the foundation of medical imaging so that you can build tools on top of that. So it requires a lot of data to train up. We’re talking mostly VLMs right now at Hoppr and in many places, visual language models that can do this sort of step towards a full understanding of imaging, a full foundation model where they take into account as many image-language pairs as they can and try to understand imaging on this deep level. So foundation models are designed to try to handle more of the complexity and context than the narrowly trained, single-task tools. And they open a lot of practical possibilities up. Instead of having to build a separate model for every task, you can fine-tune on that foundation model for a specific thing. And then, because it understands so much of imaging, you can come to it with a much smaller labeled dataset of what you’re trying to do. We’re talking, instead of in the order of millions, in the order of thousands for a specific task that is maybe important not just in a specific part of radiology, but maybe even to your medical group or to your large hospital practice. And the good thing about that is we see variability in performance of tools between practices because there’s differences in the machines, the patient population, the way people practice there. And so if you do have these foundation models that can be fine-tuned on small datasets, people can personalize their tools to their own hospital workflow and their own patient populations without that heavy of a lift. So that’s the vision. That’s the excitement.

[00:40:41] Chris St. John: Hell yeah. Okay. Before I let you go today, I want to come back to this Minority Report moment. You’ve explored VR and extended reality, also AR, in radiology. I’m curious what you think spatial computing could add to imaging.

[00:41:00] Julie Bauml: Yeah. So I actually sort of just fell into it because I was looking so much for a way to make things hands-free, mainly because an important mentor of mine had to retire early, and she no longer practices because she’s a body imager. And like me, who’s also a body imager, it’s thousands and thousands of images per case. And you have to scroll through them all manually, and her wrists just gave out. So I was looking for an ergonomic solution, any way to get my hands free. I actually injured my shoulder pretty bad myself on some of these long dictation sessions, just leaning forward and clutching the dictaphone and clutching this. So I got into it for that reason, but there’s a lot there. If you could imagine being able to replace two or three giant monitors and a desktop tower that literally weighs like 50 pounds and a dictation tool and a mouse and a keyboard setup that you need a giant wooden desk to put on, if you could replace that with something you could put in your backpack, that’s the reality now, actually. It can be done. I consult, pretty much for free, just for, I don’t know, equity or whatever, not getting paid, for a company called Luxsonic that has clearance to do that in Canada, and they’re looking for their clearance in The US. And I’ve been able to have this headset and put it on and look at cases, look at full CTs, download them, re- you know, not for clinical use because it’s not approved yet anywhere, anywhere you have a Wi-Fi connection. Starbucks, doesn’t matter. Fits in your backpack. And you can imagine, in this world where we’re so crunched on workflow and people are wanting to retire or semi-retire, giving them something where, you probably wouldn’t want to sit all day for eight hours with this on your head dictating like that, you can read a few cases here and there. You could read for hours on vacation. You could take call, and instead of having to be tied to your PACS to check stuff, let’s say you’re an IR guy, do I need to run in and fix this, or is this a surgery issue? You can put it on at dinner and check that case. So it’s just opening up the possibility. These headsets are so much better than they were, but you have to think they’re also, right now, the worst that they’re ever gonna be in the future. At some point, they are gonna be quite small. So you’re talking about the Minority of what, maybe we don’t have the ability to project it all onto the wall yet, but this would be a step in between to getting to that hands-free. And the thing about it is, it forces integration of those softwares that are not fully integrated yet. The dictation software, the EHR, and the PACS system, they basically have to be integrated. The microphone is in the headset, so I’m not clutching a microphone anymore. The dictation software has to basically be ambient and turned on and voice-activated. I’m not having to manually scroll through anything anymore. And because it’s all so integrated, bringing in AI tools is also easier.

[00:44:05] Chris St. John: Yeah. I feel like that’s such a beautiful button on our conversation, because to me, it really does touch on augmentation, literally, metaphorically, versus replacement. It’s like, it’s a whole new version to work with these softwares, with these programs, rather than our favorite ten-year-old narrative that, oh, these radiologists, they’re never gonna work.

[00:44:34] Julie Bauml: Yeah. It’s true. I’m sure someday we’re all gonna be faced with this reality that, unless we’re working directly in a physical industry like laying brick or cutting hair, that AI is coming in for some part of our jobs. But there’s a lot of benefits to that. There’s some concerns if we get it wrong.

[00:44:59] Chris St. John: Yeah. Well, alright. Julie, I think, unfortunately, that has to be our time today, but thank you so much for coming and joining us here on Rethink Imaging.

[00:45:10] Julie Bauml: Thank you.

[00:45:11] Chris St. John: Once again, this has been Dr. Julie Bauml. Julie, we’ll talk to you soon.

[00:45:15] Julie Bauml: Thanks, Chris.

[00:45:18] Chris St. John: Frame by Frame, Rethink Imaging is brought to you by Imalogix. 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 Imalogix is rethinking imaging in healthcare, visit imalogix.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.

Questions from This Episode

What is the "cognitive ceiling" in radiology?
The cognitive ceiling is the biological limit on how many high-level decisions a radiologist can make in a day. It includes both a cognitive component (decision-making load) and a visual component (how many images the cortex can process in rapid succession). Dr. Bauml argues the industry is nearly at this limit, which means efficiency tools alone won’t solve burnout.
Why did early radiology AI tools alienate radiologists?
Early tools were marketed as replacements for radiologists, with public claims that the profession would be obsolete within five years. This “point solution” approach focused on narrow diagnostic tasks rather than workflow integration, which created friction and distrust among the clinicians expected to adopt them.
What does "augmentation over replacement" mean for AI in radiology?
Augmentation means AI handles the non-doctoring tasks that drain radiologist capacity (chart searching, window alignment, measuring lesions, generating report boilerplate) so radiologists can focus on the high-level diagnostic work at the top of their license. Replacement assumes AI can do the full interpretation, which is not clinically viable at scale today.