From the RSNA, this is the Radiology Artificial Intelligence Podcast. My name is Paul Yee, and I'm a radiologist and co-host of the podcast. And my name is Ali Tajani, and I'm a radiologist and co-host of the podcast. Each month, we dive into the hottest topics in Radiology AI and talk with leading experts, thought leaders, and movers and shakers in the field. Welcome back to the Radiology AI Podcast. I am Satya Tripati, one of the new training editorial board members and an associate editor for the podcast. This is the Hot Take series. Very we discussed the ideas that move fast, challenge us options, and bring the hallway conversations of Radiology into open during a time of rapid AI-driven changes. There's no better way to launch the series than Dr. Sarpcha, many know him as Harry as well. He's an associate professor at University of Pennsylvania, and the sharp, fearless mind behind Rogue Rad on ex-formally Twitter. He brings the honesty and clarity that defines what the series aims to be. Welcome, Harry. So we will dive right into our podcast with our first segment of Rapid Question. So Harry, when you look across our synitis here, what do you see the true innovation happening in homegrown, academic tools built inside hospitals, or in winter platforms embedding into packs and EMR, or at least hoping to, or in the big tech foundation model ecosystem, like what feels most real versus more performative to see or to you? Sarpcha, thanks for the invitation. So it's all going to be performative, to some extent. It's very hard to get an idea of how vendors are truly doing, because they all exaggerate, and then they tell you that they're very busy, and they're deploying their front and center. But when you speak to the hospitals on the other hand, you don't get the same sense of adoption. So the hype is still there, but lesser than before. Now you have a scenario where we're talking about integrating information solutions, not just one disease solutions, but packages. And so the vendors are still the ones that are largely pushing AI. We're not getting that many homegrown solutions, I'm not sure why that is the case. I think that the big foundation models, they're promising, but they're yet to prove themselves as being efficacious in the workload. So it's intriguing, but still hasn't burst onto the scenes. Well, let's put it this way, if you had to choose one radiology workflow, that will be fully automated in the next five years, what would it be? And also on the flip side, which workflow do you believe will resist automation, no matter how powerful our models get? So I have to be careful that I don't confuse this with wishful thinking, so it's prediction, it's not wishful thinking, so I would want the chest x-ray to be fully automated. I don't think radiologists should be reading chest x-rays, at least in big academic centers where they do chest CTs, left right in center, and they get the chest x-ray every six hours to verify the positions of lines and tubes. But I think that's going to be the most resisted to automation. The one I think is probably a low-lying fruit. Anytime you have any measurements that come from the CT or the MR data, that can and should be automated. So for example, a workflow where the vascular structures are measured and remeasured, there would be automated, where you have measurement of ejection fraction and the cardiac MRI, that can be automated, calcium scans. So things that can easily be segmented from the images can be also automated. The major problem with that is, whilst it can be done, how do you incorporate that into the report without, or incorporating that into the report is the most important thing. The actual out of automation, the out of measuring, I think we're there right now, but it's more integrating it with the radiology information system and populating the report without the radiologist having to populate the report. And then, you know, you can just put as much as you want, all from the AI. Okay, and then we touched upon this briefly, but now many departments are debating build versus by, in the age of LM's and agentic systems. What advice would you give a radiology chair deciding whether to invest in an internal AI tool rely on vendors or partner with industry? The major drivers of vendor growth is being able to build for the algorithm use. And if they can build per click every time it's clicked, they'll make more money. So for example, if you have a lung nodule detector and every time you use that lung nodule detector, you pay a little amount to the vendor, that's the model that's going to get them dollars, but that's not the only thing that we do. We don't just look at lung nodules as a lot more in the radiology analysis. So once it's done, it's lung nodule thing, then it's kind of left dry, well, what next? I would say that the model for vendors is very much a model. The return of investment is what they're telling the radiology departments is that by measuring your nodules will make you faster, will make you burn out less. And to some extent, that's true, but it can get overpromised and it's very hard to create an ROI for that. An internal system can be seen as a investment, not just in the workflow, but also in the science, progressing the science internally. We have a rich history of doing that already in imaging, the whole field of image processing, image segmentation, creating masks from imaging, that was done without any real vendor collaboration was all internal, at least at the University of Pennsylvania was all internal. So I see no reason why you can't do the same for AI. You just have to have a very strong use case in mind. And one of them is something I keep harping on about, which is the autonomous AI with chest x-ray. What it can do is that if you can create a system internally where the AI records the lines and tubes where they are, it's not an easy task, it's a very difficult task actually. And looks for pneumothoranks and assesses the state of erasure is a more consolidation or less consolidation without worrying too much about the vernacular that we use in the chest x-ray. I think that could be very useful and that could get to the point of near autonomy and then complete autonomy for one particular area. You don't have to go through the FDA if you're not planned to commercialize it. And then there are risk models that can be used to diffuse a risk to the institution. So definitely I would go for homegrown plus or minus industry partnership, but then the question really is the industry partnership. What's the financial model? What do you get getting from the industry partnership or what you're losing? I would say that it's better to create a team of engineers and the most important thing is just to have a clear purpose what you want. A lot of times you end up with algorithms that come out of the pipeline because it's somebody's pet interest without any understanding of the implementation. So that's okay for science and that's okay for presentations, but I have a very clear idea what you need to do when it comes to developing an algorithm for clinically use. So I would add my next question is that when we talk about homegrown AI, one of the challenges a lot of the academics had us would see is that there is a research team, the research department doing and building models and then there is the clinical team and the clinical needs. And usually there's not a very well communication between the two leading to developments that do not reach the clinical space or the clinical implementation. So what would your advice to such centers and places would be who are trying to build these AI systems for clinically use and how should they communicate? So a lot of the research is dependent on grant funding and hypothesis or some sort of mini solution that they come up with to a clinical problem. Implementation is a whole different ball game. So much of the problem occurs because the research team thinks something is a good idea because it's very interesting from a computational point of view, from a research point of view. And when you throw that idea, the clinicians, they look at it and say, well, no, what uses this? An example is prognostication, prediction is a whole host of AI driven research predictions, research you had during predictions for various things. And that's not clinically useful because we kind of want to know what is happening now. What may happen in five years from now isn't a useful thing to know now. It may be useful for the patient, although I don't know that would be that useful. But if you have a 20% chance versus a 10% chance, what will you do differently? What is clinically useful can be bland on the research side. So you have to create a clinical team with a particular project in mind. It has to be the same way as you have same kind of format as in the corporate world where you have a particular project, you have a project leader, and then you have others that have very much assigned jobs. So it's not just a communications challenge, it's more of a leadership management challenge. But you have to identify it specific. This is what I want. Now can you enable it? Another example would be using radioenomics in clinical reports to guide therapy. So first you have to, of course, develop that, that would be an in-house development. There's not much point of ended doing that because there's going to be so much difference between one institution and another. And then you have, you use that once you've developed that through the clinical research team, roll it out and give that information to the oncologist so they can decide what they want to do with data coming from. So in the end, you've got to make it actionable, but you also want action. So with information, you have to, you first have information, then you have, is this information actionable, then it doesn't actually lead to action, those three things. So one of the interest for both research and the Wenders, which they have been advertising are these co-piles, our L and co-piles that could be inside of PACS or EMR system. From your perspective, what does a useful co-pilot look in a radiology facility and what would actually make sense to have and what is just another button on the screen? The most important thing a co-pilot needs to do is to learn to shut up and be invisible. So the problem with co-pilots is that they can quickly become a monkey on your shoulder, which is why CAD for mammography was dismissed. So I'm not a huge fan of co-pilots, either take the damn thing entirely or get out. This co-piloty thing is just an useless, but I can be persuaded otherwise. But just if, for example, I'm reading a, you know, a chest CT, a co-pilot points out all the lung nodules great. Thanks for doing that. And now what is happening, what's going to happen is, either I'm going to ignore it or I'm going to depend on it, but I can't halfway ignore it or halfway depend on it. That is, it's a quantum impossibility. And if you try and do that, then that creates an additional burden on your mind where you're looking at it that you say, "Oh, let's look at the co-pilot." "Oh, but the co-pilot's saying this." Now something might say, well, use the co-pilot when you're unsure, fair enough, but when you start using the co-pilot when you're unsure, then you can drift to a situation where you completely neglected, or you, I think, are willing to be proven or wrong. But as or now, I think the most important thing with the co-piloty, it needs to remain quiet. And its recommendations need to be just obliterated in the ether, because once it starts being recorded, that creates a legal risk. But on the same line of thoughts of having more monkeys on your shoulder, when you walk down the earth's navel or in general, look at the vendors, what are some messaging that sounds super similar, which is more overrated or over-promised ideas you have seen so far related to these foundation models, co-pilots, autonoms, reporting, so on. I don't know who they're appealing to, but I imagine they're not appealing to me. I don't think that they're thinking, "Let's appeal to this guy." So, I don't know, I mean, do radiologists want a co-pilot? There's some companies that embed reports, so if you describe all the findings, it'll give you an impression. I find that useless. I find that utterly useless, but people swear by it. So there's obviously going to be heterogeneity in preferences amongst the radiologists, which is essentially what's going to determine the market dynamics. Obviously, you'll need a certain group of radiologists to go in a certain direction for a particular product to succeed. So I wouldn't generalize my opinions to radiologists at large, but the messaging started off with, "We will replace you, and imaging is as easy as cats and dogs, too. We will sit by you, which is really annoying." Like, "Get out of my car." You know, I'll take your car, I'll just sit in the passenger seat and go wherever you go. So if I have to hold a steering or a Tesla vehicle, I might as well drive it as well, if it's not driving on its own. If you're perpetually paranoid about letting go of the steering wheel, and you're like, "Oh, this is going to go off in a ditch." Then you may as well just grab the steering wheel, because if you forget how to drive and it does go in a ditch, then you won't be able to do anything. So again, either take my entire work or leave me alone, don't give me this, I'm going to jump in the box with you. Well, move a little positive, so on the other side, not being positive, it's very positive. But I'd be very optimistic. I want AI to succeed in the right manner, like I wanted to take our mundane work promise. It's simply not doing that. Yes, and I feel the field is not quite understanding the role of AI in how it should be autonomous in a way that it helps and not trying to replace it, should be an addition to limit the workflow. And as what we have been doing in workflow optimization for years, even before AI has come into play, I think it's the same strategy, but the marketing has changed quite a lot. So on the other hand, what is one undiscussed or quiet innovation that you see? Either a tool, workflow, idea, research direction that you think could have a major impact, but isn't getting enough attention. I think radio mix can have a major impact, because the idea of radio mix, it gives information about biology that goes beyond what our eyes can see. I mean, of course, the way it would go is that you have information about a biology, the phenotype, but also there's genetic information as well. Now that has potential, because if you've got all these very expensive drugs and you find a subset that it doesn't work at all, you could exclude that subset. So I think that's underutilised tremendously. I think I like the algorithms that reduce scan time. I think billing it as a, you could get more patience through the door is fine, but then you're kind of, again, appealing to the administrator. You could use it for artifact reduction, so I think that's a potential. So speaking of stakeholders, collaboration is becoming more complex, so academics, vendors, EHR companies, cloud providers, and startups all want to stake in this AI pipeline. So what does a successful collaboration actually look like in 2025 going into 2026? And what does a field collaboration look like on the other hand? Well, I think the successful collaboration is one where the department works with the vendor. It's richer as a result of the work, and the department gets smarter as a result of the work. But then you have to have a clear idea, and you have to clear a financial model for that. A failed collaboration is when you, when a vendor decides to push the algorithm prematurely, get a couple of publications here and there making everybody happy, but nothing comes from the implementation. You obviously need informatic gurus to integrate algorithms of the workplace. The problem I see right now is that there are the academic departments that once were the pioneers of CT&MRI pushing the technology, both because they had a sense of what they wanted to achieve, and also because there was a scientific experimentation. I'm not seeing much coming from them. They have to drive it. The vendors are the cart, the vendors are not the horse, so they have to drive it and they have to tell the vendors what they want. And in fact, what we have now is just a suite of algorithms doing a bunch of things of variable usefulness. So you have perfectly taken me to my next question, which is, is the innovation now has been led by resource and vendors, and where does that place smaller practices and smaller academics institution, and we're just talking about US or not even talking about our global clinics, who are completely ignored most of the time when we talk about these AI innovations. So small practices, AI is the least of their concern. Like if you go to a small rural practice with staffing issues, with delivery issues, and the last thing they want to do is hear about AI, and it's the classic Mari Antoinette problem. We don't have radiologists. Oh, what about AI? So the adoption is really going to be in these big places, and in these big places, the adoption is going to come when there are solutions that can be scaled. Algorithms can only be scaled if they're solving something that needs to be solved. You and I both share interest in health equity, health policy, tariff line. So in a scenario where AI does become the standard of practice or is the quality of care, do you think it becomes a health equity issue that now a smaller practice issue also be moving towards AI to provide the best care to the patients and not just a smaller practices, but also expanding it to our global health initiatives that we work on. Well, the moment AI is featuring in places which don't have radiologists, so you already have a health equity issue where the poorer and the underserved in the world are getting AI. So what you're talking about is the opposite where the richer people get in the air and the poorer aren't. Yeah, I think you'll always have that with technology. The most important thing is to make sure that technology is not so life-saving. If technology is very life-saving, then it should be available to everybody. I don't think we're there yet with the AI at all. So moving to the last question of our rapid question section. Looking ahead, how should radiology department prepare their workforce? So including trainees, residents, faculty for a future where these agentic systems or co-pilot system automate task, trianche cases, draft reports, potentially, and cardinate workflow steps end to end. So what is the new skill or the mindset that should be essential? Well, you're going to have to go back to your driverless car analogy. What's a new skill required to drive a driverless car? And if the answer is not driving, why are you there? Even so far as you're going to be required to furnish your opinion in a certain number of cases, you still would have to go through the same, maybe not the same vigor, but the same training in order to do that. Now, it depends on how what the end product is. If AI is measuring liver metastases, then of course the training doesn't need to do that. If AI is measuring it autonomously, the training doesn't need to do that, but if AI is measuring it, they're training needs to check the work, they're training needs to know how to do that. So it all depends on the level of involvement. There is definitely scope for moving to the old model where the radiologists was very much of a consultant. So being able to look at multiple scans, read the patient's history and be a consultant and give a differential or give it back and forth with the referring physician, something that doesn't happen at all, something that can happen should happen. So it depends on the level of, you can't prepare for a scenario that hasn't happened. It has to be piecemeal and you have to be flexible to adapt to whatever the new situation is. So now to our last segment, three predictions, three fears, where you will give me three bold predictions for AI, foundation models and elements in radiology and then three fears or cautionary signals. The first thing is that you will get a scenario where AI does a lot of the measuring. People are beginning to realize that that is a low-lying fruit that doesn't have the same appeal to Silicon Valley like replacing radiologists, but that is a, so I think the measurement plague, the measurement epidemic will be resolved, hopefully, as prediction number one, prediction number two is that there will be a lot more patients looking at chat GPT for diagnosis, meaning of radiologists reports and even putting their x-rays into AI. So you will start getting conflicting reports with AI and the work of a radiologist will increase in order to resolve these discrepancies, prediction number two, prediction number three is that homegrown algorithms will not be exploding in the next three years, it's still a mindset amongst the academic medical centers, all right, how you said fear. Well, the fears are really kind of doomsday fears aren't they? Like AI will start an internal war, AI will, it's like Elon Musk's fear about AI, AI will be civilizationally dangerous. I think one fear I have is that I have a fear that AI will be over-regulated, that will stifle innovation, and I have a fear that AI will be under-regulated, so you just have a whole lot of algorithms just popping up here and there. And so you and I both are going to have over-regulation and under-regulation, so I've used two of my three cards in that, so I have a last one. I think radiology is going to go through this supply demand problem because if AI marginally improves that it starts making news, then there will be this whole avalanche of predictions about AI. Radiology is not existing, there will be a dramatic reduction in the radiology and then there will be a shortage of radiologists, even worse than what it is right now. So I don't think there's much you can do about that because if you try and overcompose for that, you could get a surplus problem. So the workforce is going to be a big problem in radiology. There could be dramatic changes in interest based on how the media portrays AI's success. That brings us to the end of the episode. Thank you, Harry, for being here, bringing your trademark honesty to this discussion. I'm grateful you agreed upon doing this series with me and thank you to all our listeners. Thank you for the invitation. Good luck with the series. We hope you've enjoyed this episode of the Radiology Artificial Intelligence Podcast. Shout out to Yuri Aishin Chiesin for our podcast music and to the RSNAY Podcast staff. Email us at
[email protected] anytime with questions, feedback or suggestions.
PodcastTranscript AI