← 返回

2026.10.06 | TB | MedTech | Medical Imaging AI: Fujifilm & Lunit Adoption

原标题：Medical Imaging – Customer Perspective on Fujifilm & Lunit AI Adoption Progress

#medtech

Third Bridge · 2026-10-06

Analyst-led • 18 Sep 2026

Companies: 4901 JP, A328130 KR, RDNT US, GE US

Robert Braun (RB) COO at University Radiology Group LLC

Michelle Lee (ML) — Third Bridge Sector Analyst

Agenda

Key Insights

Introduction

ML [00:00:06]: Welcome to Third Bridge's Interview entitled Medical Imaging – Customer Perspective on Fujifilm & Lunit AI Adoption Progress. Hi, I'm Michelle Lee, and I'm joined today by Mr Robert Braun.

Before we start today's Interview, please state I agree or I disagree to the following statement: You understand the definition of material non-public information and agree not to disclose any such information, or any other information which is confidential, during this Interview.

RB [00:00:42]: I agree.

ML [00:00:44]: Could you please begin with a brief introduction to your background?

RB [00:00:50]: Absolutely. I have been in the radiology diagnostic imaging or nuclear medicine space for roughly 23 years. I'm also a clinician currently, and probably for roughly the last decade-plus, I've been in an executive role as well, that role being Chief Operating Officer. Today, I'm in that role for a very large institution, 200-plus radiologists covering and owning about 40-plus locations. We see two million-plus studies annually, both in the inpatient and the outpatient setting as well, sub-specialised, meaning we read within our lane, how I usually put it, where the majority of the studies are read by fellowship-trained radiologists, whether it be in MSK radiology, neuroradiology, women's imaging, middle-market company and a very pronounced IT stack of platforms that we utilise from both a business case, but also a clinical utilisation use case as well.

Questions

00:02:10] What do you estimate to be the TAM and CAGR for global medical AI over the next five years?

RB [00:02:28]: Over the next five years, the market size is considerable today, and that considerable is not in the millions, but in the billions. I would say roughly, and specifically in the radiology area, will be probably between USD 4bn and USD 5bn in the next five years. I could see that growth exploding to at least USD 13bn, which is significant. It is a very competitive space. There are many products out in the market competing broad use cases. Radiology is a very broad segment as well, but that presents opportunity for a lot of AI companies because of the diversity that is required in medical imaging. The detail behind that, everything from assisting physicians in terms of triaging cases, which is a common use case, would definitely include workflow optimisation, which is something significant due to the radiologist labour shortage, which is very significant. That's why AI is getting an additional boost, so to speak. The third is the assistance in clinical identification, meaning it will assist the radiologists, whether it's a mammography study, PET study or CT study in assisting them identifying potential pathology.

00:04:13] What is the ratio of each use segment you mentioned, such as triage or workflow optimisation and studying diagnostics in your hospital?

RB [00:04:34]: I would say right now, I'll put it in those three categories, triage, workflow optimisation and clinical decision assistance. I would say reasonably, it is probably around 40% triage, and the remaining split, 30/30 workflow optimisation. I think workflow optimisation, in my opinion, is starting to get a little more momentum. It sounds a little less exciting and maybe it is, but I think why there's more momentum in that area specifically is because of that radiologist shortage, which we know as an industry is not coming to an end anytime soon, which means YoY, the request for more imaging is growing, but the growth of radiologists is stagnant. It's depleting. There are not enough radiologists coming out of their fellowship or residency to replace those radiologists who are retiring.

00:05:46] Besides the shortage of radiologists, what primary regulatory and policy drivers are accelerating the adoption of medical AI? What are the major institutional barriers hindering its market entry?

RB [00:06:08]: I'll begin with the biggest barriers from a high level, which would be, at least based in the US, FDA clearance. This is something that is a rigid process for anything within the medical space for US-based providers. When you throw in AI, it really goes through a rigorous process on many levels, obviously, in radiology, which could delay implementation. Many people are using AI, but not from a use case. They're just using it to validate it at this point, but it's not been approved at all. You might work with a company that has 10 modules, but only three of them might be approved for real clinical use within the US itself. The rest are still under a clinical trial or on review by the FDA. That's the US.

It's going to be the same thing in the European market. The European Union will have a similar process in place that mirrors what the FDA does. I think one of the challenges there is there are so many players within the market that it's taken longer for products to get that actual clearance itself. I would also say there are some internal pressures in the US by many lobbying groups to move these AI platforms along because of the shortage of radiologists in the marketplace. One thing I would say is that since there are so many players, what I've learned very quickly is the more AI is used, the better it gets. I can hear that in a lot of news because of machine learning of sorts. The more data that's pushed through it, the algorithms get better. I think we as a medical community are seeing more and more clinical evidence that substantiates the need for using AI within the space.

00:08:38] Who are the top global or US players in the medical AI space? We were discussing Lunit. Among the top players, what are Lunit's core technological and commercial differentiators?

RB [00:08:59]: What's interesting about Lunit as opposed to other providers in the space is they represent a lot of the similarities in terms of what other providers are offering. Again, triage, they don't really offer workflow optimisation from my experience. Their little claim to fame, and it's not approved at least to my knowledge in the US, is more of their oncology module that is still under review in the US itself, which I think that could be a differentiator. Why I think it could be a differentiator? In the world of medicine, particularly radiology, some of the major areas of imaging are going to be things such as cardiovascular imaging, neurological imaging and cancer imaging, which is very broad. This is an area that you don't see as much focus, I think, my personal opinion on. Most of the AI companies are focusing a lot on the triaging component or broadly focusing on identification of disease. I think Lunit is probably trying to get more granular, specifically in the oncological space of it.

00:10:28] Who are Lunit's competitors?

RB [00:10:30]: The major competitors, and there's no shortage of them, are going to be companies like Aidoc, Rad AI, Annalise, ScreenPoint and DeepHealth would be the ones off the top of my head that would be jumping out for sure.

00:11:03] What do you think is the most important factor for a player to survive and be successful in the medical AI space?

RB [00:11:16]: To survive, I think clearly the number one thing is validation of their data, meaning how good is the accuracy. At the end of the day, whatever the AI solution is trying to solve, if it's not accurate, nobody is going to buy it or implement it within their space. Accuracy is going to be key. Very close to accuracy, though, is going to be speed of how quickly it can be done. When I say speed, it means a radiologist is under the gun, under pressure to read as efficiently as possible. If you put on many layers of multiple AI modules, they constantly have to click through and open different windows to use the AI, which is very inefficient. You want a platform, best-world scenario, that it's comprehensive, that it can do everything from mammography, triaging, oncological imaging, and workflow optimisation, all on one platform that is seamless without that radiologist having to be overwhelmed with booting up and adding different passwords, etc, to utilise the platform. You basically want the radiologists to be able to dictate their study seamlessly with AI working in the background to facilitate them reading that study itself.

00:13:01] Some global medical device OEMs are trying to internalise AI software into their hardware. Is that an opportunity or a threat for independent healthcare AI companies such as Lunit?

RB [00:13:43]: Absolutely. I'm sorry. It is and it's not only a threat, it's already happening. What I mean by that is speaking to some major vendors in the radiology space, public knowledge, this is the thing that we use already, companies like GE, Siemens or Philips, common OEMs within this space. What they have done in their next generation of technology, and that technology being the imaging component, the MRI machine itself, that CT unit itself, they've built in algorithms to, in many cases, speed up how quickly things are imaged, but that's a stepping stone. They're using those same algorithms to enhance images to make the radiologist's job, again, easier by identifying pathology or quickly again, triaging cases and things like that. They're looking to integrate so that you can become a one-stop shop. I don't think, in my personal opinion, they're doing it to push out third-party companies. What they're doing is to be competitive in their own OEM market because their competitors are doing it. I think the third-party companies are just side collateral. What I see as an opportunity for the third-party companies is they have spent, in many cases, years in developing their AI algorithm that is fantastic.

You participate and work with those OEMs to integrate it into the next generation. We've seen that to a certain degree with PACS systems, which have enhanced AI. Again, public knowledge, GE, a big player within the radiology space, they made a purchase 6-8 months ago of Intelerad because they were a cloud-based solution with AI enabled. They weren't going to spend years trying to reinvent the wheel. What do they do? They go out in the market, find who has a lot of momentum and they buy them out, and now they'll integrate it, I am sure, into their existing platform and get to market a lot quicker.

00:16:12] Do you think that Lunit's strategy of getting into Fujifilm's hardware is clever in terms of surviving in this market?

RB [00:16:24]: Absolutely. I've been doing this a long time, and what we've seen is an explosion of organisations, including mine, jumping toward all new AI. You realise very quickly it's overwhelming by bringing on multiple modules, it makes it more complex for the radiologists. If you can find that one platform that has best-in-class in the majority of places, one, it will more streamline the workflow for that radiologist, which in theory will make them more productive. From a business perspective, more economically favourable, meaning that I may not need 7-8 vendors to do all the AI I need. If I can get one vendor to do it, I may be paying per module, but usually, when you work with any of these AI vendors, the more modules you buy, you basically get discount pricing. If they had five modules and you're using one, but decide to add another module, they want to retain you as a customer. They're going to give you aggressive pricing to go on to that next module.

00:17:49] Let's focus on studying diagnostics and clinical decisions. Beyond reducing turnaround time, what are the primary incentives for hospitals to adopt AI imaging diagnostic systems?

RB [00:18:09]: Besides the actual reading time itself, what we look at is, and then, I said, "Well, most companies look at the internal metrics when they apply AI or when a radiologist section, let's say, is not applying AI-enabled systems." The basic one is how long the turnaround time of a new study is. That's the gold standard out there. We're also looking at the number of radiologists we need per shift to cover the number of cases. If I'm a company that is private, let's say I have 50 partners, but I employ other radiologists. If I can employ less radiologists because my system becomes more productive and my process, that in the long run is going to save me overhead. I may not be generating a per-click fee by using AI. What AI is doing is allowing me to reduce my need to spend on hiring 2-3 additional radiologists, which could easily cost me USD 2m-3m annually. If I paid for AI and it's costing me USD 200,000-300,000, but it's saving me USD 2m-3m, that's the other metric, what is my return on investment for my AI investment.

00:19:51] The AI image-reading reimbursement is quite limited currently. How much does cost saving justify large-scale capital investment or initial investment into a comprehensive AI system?

RB [00:20:12]: Your comment is true. There's not a lot of momentum in the space for putting a system in, and obviously, in our US system, whether it be a private payer or CMS, which is Medicare, right now there are very few opportunities for getting a per-click fee for AI. I think that momentum is starting to slowly grow. There are a couple of opportunities that exist, but ultimately, I think it's going to backfire. What I mean by backfire is Medicare, in their ultimate wisdom, in my opinion, they're going to say, "Hey, AI, why are we paying you for this? It's making you more efficient. We should actually be paying you less because you're doing more studies per hour now, and now we're paying out more money. Why are we paying you for AI? AI is giving you the opportunity to actually generate additional income itself." How I look at it, they're going to look at it or spin it as double dipping itself. I don't think there's a lot of pressure on that right now just because of the massive shortage, but I still don't believe there's going to be a lot of opportunity for cost savings or for getting reimbursed for AI itself.

For any company who's trying to build their product out, after using AI for 6-7 years-plus now, what we see, and I think the industry, it's plateaued, meaning that the metrics that we use to validate these products, whether it be Aidoc, Rad AI, etc, the value proposition levelled off. After you have it in place for 1.5-2 years, you're not seeing a continuing decline, meaning that if I've seen an improvement in turnaround time of 10%, three years later, I'm not seeing it at 50%. I think most people are seeing a max improvement of about 30%. That's being very efficient and utilising AI to its full extent. In some extent, it's plateaued in my opinion.

00:22:58] Besides saving labour costs via hiring fewer doctors, as you mentioned, how do you evaluate the contribution of a medical AI solution in operational cost reduction for larger hospitals vs small-sized clinics?

RB [00:23:20]: Good question. What I would say is outcomes. What I mean by outcomes is, I'll give you a real-world scenario, let's say, Aidoc. Aidoc, like many of the companies out there, has a triage. If an individual is coming in because they're having a stroke. Instead of sitting in the waiting room for 30 minutes, they scan the patient, and that scan could sit on that theoretical pile for radiologists to read. Let's say it was case number 15 to be read. Case number 15, you could be sitting there for an hour. A lot could happen to a patient in one hour without medical intervention. In the case of Aidoc, what it does is the patient was scanned. As soon as the scan is done, before that technologist, emergency room physician or radiologist even opens up the study, the algorithm literally looks at it and says, "I think this patient is having a stroke." It takes it from case being read number 15 to number one, tells the radiologist, "Stop what you're doing, look at this case." They open it up and they either say, "No, AI was wrong, this is a congenital anomaly," or "Yes, this is a stroke." They dictate it out, they immediately send it back to the emergency room physician and they act accordingly.

What's the benefit of that? Not just because the patient didn't have to wait as long. The sooner you could act traditionally on a patient in that acute phase is going to have a better outcome. That is meaningful, obviously, for the patient and for an insurance company because now a patient who was able to be treated much more quickly, and they have a better outcome. They may not be disabled, which would have incurred more medical expenses. If you're able to catch that medical emergency a lot earlier, you're actually not only helping the patient, but you're also reducing financial burden to an insurance company as well. That is one of the outcomes, and it is definitely measurable.

00:25:45] What is the current pricing or billing model adopted for Lunit within your network? How has it changed over time?

RB [00:26:07]: I would say that the pricing model, they're competitive in the space with the others. They are priced competitively to their peers. The only thing that I would say, and it's not just specific to Lunit, it's to all the providers and competitors in the space, is they are going really to this subscription, software-as-a-service model, where basically you have an annual subscription. You often will get a discount if you lock into a multi-year agreement as well. It's also tiered where they offer tiered pricing, meaning that it's like a per-use case. You might have a maintenance fee that you're being charged, but it also could be a per-click. If you don't use the system, you don't get charged for it, but if you do use it, you do get charged for it. An example would be, let's say you're using a system for nodule management. If I have 200 radiologists and some decide, "I like the system and it makes me efficient, etc," they use it. We are only going to get a charge not per radiologist, but per case that they use it for. An AI company like Lunit wants the radiologists to be engaged and use it on every study that they potentially could use it because they're going to get a per-click fee.

We like that model. Why? Just because we're not being charged for something that maybe 20-30% of the group have decided they don't want to use. In my situation, I'm a private group, so it's a little more liberal. If you're in an only academic institution where you're employed, your employer may say, "Hey, you have to use the system. We're paying for it, you need to use it." I have been in organisations, to be honest with you, where they do it after the fact, and that doesn't help you. They just do it to validate some findings sometimes, but it's not what really most organisations want them to do. They want to use it as a partner system so that, again, it makes them more productive as they're reading the studies.

00:28:52] I'm unsure that Lunit is increasing prices. As you mentioned, it's a competitive market, so there could be a discount. If the company is locked in, it can increase its annual subscription price in line with CPI increases. What are your thoughts?

RB [00:29:16]: The other thing that you can talk about them is they can go not just to the group, but to the patients where, since insurance is not reimbursing it. I don't say this is a negative way. Basically, they are allowing a patient to have it done. The patient takes the study to Lunit itself and it's evaluated, it's run through their system, and they basically bypass the insurance, which you're allowed to, and pay out of pocket for the service. We've seen that with other areas, companies in the area of cardiac AI like HeartFlow or Cleerly, patients are driving this. They find out, insurance doesn't pay this, or they may pay it but not in your instance, but you want it to be done as an added safeguard. It's non-invasive, things of that nature. If you already have the study, you're basically sending your images to a company who will use the AI algorithm to analyse it to validate whatever interpretation you had received. Obviously, if it differs, it gives you an opportunity to go back to your physician and say, "Hey, maybe additional testing is required."

00:30:44] Are the increases in workflow efficiency or resource saving from Lunit AI in diagnostics tangible or measurable at the operational level?

RB [00:30:59]: Yes, I would say definitely. In general, using the system has definitely improved the overall workflow. Again, it's in line with my figure of about one-third, let's call it round numbers, one-third of improved productivity of the radiologists. The best opportunity that Lunit has, and it's not just theirs, where AI began is in two broad areas. It began in mammography and in lung nodules, two areas that many AI companies, especially early on, focused their energies on, high-volume studies, high-risk studies, things that are usually looking for usually cancer. This is the area that I think there's continued momentum. I would still say I haven't seen much higher than about one-third of improved productivity of the radiologists and mammographers, which are those radiologists who are reading mammography studies. They're really under the gun. When we talked about radiology shortages, this is probably one of the hardest-hit subspecialties within the space. This is an area for AI to shine, meaning that if you can give me a product that is going to halve my time of reading studies, they're going to be a clear winner. Everybody in this space, including Lunit, I would say is about the same.

When they implement it and you have a year's worth of use, they're at the 30% mark, maybe a little north of that, but I haven't seen many companies move the needle much higher. If you can get a company, including Lunit, if they're able to fine-tune their products that can get me to reduce my reading time by 50%, that will be a home run because, like I said, that's a major gap that the US is experiencing. In any group, any radiology group, any medical centre that I speak to, the hardest-hit area is mammography.

00:33:25] What was the decisive factor for your organisation choosing Lunit over its competitors? What is the KPI, such as reduced reading time or price competitiveness?

RB [00:33:43]: I think the biggest is accuracy. At the end of the day, as I mentioned earlier in our conversation, it is paramount. You don't want a product that is faster but less accurate. That is not going to help the patient, and definitely not going to help the reputation of your group. Accuracy, I think, was the driving factor itself. A close second is the ability to improve the turnaround time or productivity of that individual radiologist itself. I would say the biggest area with them is that they've also plateaued. It's a great system, cost-effective, but again, kind of plateaued. The one thing I will say about it that's kind of a niche area in breast imaging itself is the concern around what they call dense tissue breast. In this particular area, this is where they try to point out that the system is much better. If a patient has very dense breasts, you're seeing a much higher accuracy in detecting cancers. That's a win probably for them. I think that's something they highlight when they're trying to attract new customers as well. That's an area I think they should be continuing to highlight as well.

00:35:34] You mentioned accuracy. What false-positive or false-negative rate is considered acceptable?

RB [00:35:48]: I think the number in terms of what they call the false-positive rate, it's a high number, let's call it a 90-95% reduction in what we call false-positive marks. That's compared to these older CAD systems that were in place.

00:36:27] How has the accuracy of final diagnostics changed between pre-AI adoption and post-AI adoption? Have you seen any trends or changes?

RB [00:36:51]: When you look at mammography, there's always a disclaimer on mammography notes that traditional mammography, whether it's 3D or not, has limitations, period. That means that even a well-trained radiologist could miss on a low end probably 13% of cancers when they have a screening mammography itself. That increases when a woman has, again, a dense breast. With adding AI, you probably reduce that. The accuracy has now improved that naked-eye issue. It easily has brought the accuracy close to 95%.

00:37:48] If you have a normal PACS [picture archiving and communication system] or EMR [electronic medical record] system and you're trying to integrate the Lunit solution into those systems, is the integration seamless enough to avoid any clinical delays, or does it take time and effort?

RB [00:38:30]: What I would say is no. Most of these systems, including Lunit, are built to be vendor-agnostic, which they should be because there are hundreds of PACS systems within the space. If a vendor like Lunit limited their ability to integrate into a PACS system, they're basically cutting themselves out of a huge part of the market itself, and the PACS system market is changing very quickly. Not that people change PACS systems over and over, but as new players enter the market, things of that nature, they may adopt a new system. I would say they work well with what I've heard with most PACS systems. There are always a couple of bumps when you add in, even including Lunit, integrating it and layering it upon your PACS system. Once it's done, I would say that it's consistent with other systems in the space as well. One thing I would mention is when you think of the competitors in the space, the competitors in terms of OEMs is more in the PACS area because PACS systems are the ones who are building in algorithms so that you don't need a third-party solution. An example would be Fuji PACS. Fuji PACS has multiple AI solutions. They offer what they call Workflow Orchestrator, which is an optimisation tool itself.

These are things that are probably more of a threat to these independent third-party companies because Lunit is not a PACS system. Fuji is. It's easier to create an AI algorithm in an existing PACS system than it is for an independent AI solution to become a PACS system. I think the trend, I put it this way, if I'm a player in the market, I'd rather be the PACS player building AI into my solution than being the third-party solution of AI trying to get implemented into a company who uses McKesson PACS or GE PACS, because eventually, those OEMs are going to find a solution. As a customer, if their solutions are equal if not better, why would I use a third party? When you use a third party, now I'm going to incur the cost of integration, potential gaps with that integration in our world of data exchange. If you miss some data, that could cause errors. When you're using one system, the chance for error is going to be reduced. Again, if you can have a native system, that is something that I think most organisations would almost prefer vs using a third party.

00:42:08] Each hospital utilises equipment from different manufacturers and operates under different radiation-dose conditions. Does Lunit demonstrate consistent performance across diverse environments? Is any automatic adjustment required?

RB [00:42:36]: If I understand your question correctly, in the actual imaging equipment itself, there are these AI algorithms in the OEM machines that basically, from a safety perspective and patient care perspective, have built-in algorithms for the radiation. Not all imaging is radiation-producing like ultrasound, and let me clarify that, not all of it is ionising radiation, which is the most serious like X-rays and things of that nature. These algorithms are built in, they call it phototiming. Phototiming basically only uses the amount of radiation needed to produce a quality image. It doesn't overexpose a patient. They basically have a dose reduction algorithm without any reduction in the quality of the image. That's another advantage where a vendor, if it's a native solution, that GE of the world, the Siemens of the world, they have the capability of doing that because it's their hardware. Now it's basically building software to make the product better, and in this case, making a reduction in radiation exposure, which is huge. From my experience, just a month or two ago, we put in a new CT unit. The CT unit has a significant dose reduction, but the image quality is better than the old unit. It's used by some type of upgraded software. I don't need to use a third-party company to do that. It's already built into the OEM.

00:44:41] You discussed performance improvement post-AI adoption. Has the daily reading capacity increased significantly? Has the decreased fatigue per scan improved the radiologist or healthcare professional's overall quality of work?

RB [00:45:17]: Yes, and I'm going to be biased with this. I think it's a bias that we see in the industry. I don't want to paint this with a broad brush here, but I think we've seen earlier adoption of AI, including Lunit, by radiologists who are younger. When I say that because they've gone through training, more recent training, their residency, fellowship, etc, where they were immediately exposed to this from the first day. If you have a radiologist who's been in the space for 25-30 years, they went through their residency training or fellowship training without the value of AI. They've basically been working for 15-20 years, and now it's an interruption. If you learn how to dictate a case and you've been doing it the same way for 15 years and all of a sudden somebody comes along and goes, "Let's layer this type of software on there, it's going to change your workflow," that could be disruptive if you have a rhythm after 20 years. The radiologists who basically, again, were trained as residents using the system from day one, it's ingrained into them right away of the value proposition of AI.

I'll actually say this, this is my personal opinion, it could make that radiologist in training a little more dependent on software vs their own personal expertise because they're almost over-utilising the software to give them the diagnosis, so that could be a slippery slope. Again, if AI ultimately becomes more accurate, that might be the avenue. I don't think it's going to happen. AI is not replacing the radiologists regardless of all the stories you hear out there, especially in the US. It's much of a political component and maybe more political than reality of what AI can do. At the end of the day, that's the scenario itself. I would say, overall, it has improved the productivity of the radiologists, made them more efficient. I'll actually say this, more happy. When I say more happy, it is because there's less stress on that radiologist to read so many cases per day. They don't feel like they're under the gun as much.

00:48:01] The new, younger radiologists are becoming too dependent on AI systems, and technology is becoming more advanced and more reliable. Do you expect the radiologist's primary role in image reading to shift entirely to that of reviewers and final decision makers?

RB [00:48:38]: I would say it could lean in that territory, not in the very short term, but maybe in the longer 10-plus year time on what I would say the low-tech side at least. The low-tech side, when you think of radiology, think of it in two broader categories. Low-tech is things like X-ray and ultrasound. High-tech tends to be the complex cross-sectional imaging like CT, MRI and things of that area. AI assists a radiologist or a physician in identifying a fracture in a basic X-ray bone is huge because when you think of radiology also, the number of X-rays, like chest X-rays, etc, is probably the first type of case a radiologist learns how to read, a chest X-ray. It's so common, both in the acute care setting and in the outpatient care setting. Since it's high volume, it represents a huge opportunity to improve productivity. I would say you're already seeing a lot of these algorithms out there that focus on chest X-ray interpretation or basic X-ray, like bone X-ray interpretation, because it is, I want to say, less risky.

Obviously, you can have lung cancer and things of that nature, but I think because of the sheer volume of cases that go through the algorithm, as I mentioned before, the more data it does and processes it, the better it gets. I think some of the algorithms out there for this are very good. In that case, I could see organisations saying, "Oh." I'm not going to say a country, but I've already heard about a country who in rural areas are dependent on AI looking at chest X-rays, where the only time a radiologist looks at it is if AI pinpoints something and says, "Hey, there's a positive here." That's routed to the radiologist to re-review to see if it is really something of a pathological nature itself. That's great because if you're in an area that has really no radiologists and you're seeing hundreds of thousands of patients a day, AI is the value and only makes the radiologists look at cases that warrant a human actually looking at the case.

00:51:25] What percentage of your radiologists have proactively accepted or are using the AI imaging systems currently?

RB [00:51:39]: I would say about 80%.

ML [00:51:43]: How about the remaining 20% of radiologists that are refusing to use it?

RB [00:51:51]: The rest, I would say, are physicians who are, because of our own making, limited in nature of what studies they read. They're toward the end of their career, and they only want to read certain studies where AI really is not needed. They're basically looking to retire. Trying to introduce a new process into it by the time they actually get it to a point where it's of value to them and making them productive, they probably have been retired. It was not cost-effective in terms of time of introducing it to them if there was no interest in learning it because within two years, they were retiring anyway. A year just to implement, get them trained and become more efficient, and they retire. At the end of the day, we've wasted time and energies in educating somebody who is not going to use it in the long term.

00:52:58] From the users' perspective, for Lunit to become more popular or more proactively used, what are the most urgent UX or technical challenges that should be addressed?

RB [00:53:28]: I think specifically for Lunit, you mean in terms of the modules that are not approved in the US?

ML [00:53:45]: Yes.

RB [00:53:45]: One, I would say some of it is not their own doing. It means the backlog and timing with the FDA and that level of bureaucracy when there are competing systems being reviewed at the same time. That's one. Part two of that is their clinical trials part of it. There's only so much data out there in centres who are trialling the product. The less access you have of organisations piloting your system to get that FDA clearance, the longer it's going to take. I think there are other bigger players already have a big foothold in the US, like Aidoc or Rad AI, who have that momentum. They're able to evolve their product a lot quicker because some things may not be FDA-approved, but organisations may be doing it off-label, just piloting it in a sandbox of sorts. I think that's probably one of their challenges. I think the biggest challenge is that there's a lot of competition in the market and groups are quickly overwhelmed with options. If I go to a national meeting, like RSNA, Radiological Society of North America, there are literally hundreds of these vendors in the space.

Realistically, I can only look at so many. After a while, it starts looking all the same because it's a different algorithm, and there are nuances between them, but unless there is a clear differentiator in the market and I'm already using a system, I'm not going to switch out. I'm not going to say who they were, I was using one system. It was ultimately pulled out after about a six-month trial because it was not easy to learn. There were also issues on the technology side even. We said, "You know what, if it takes this long to evolve and adapt, in the long run, it's not worth our energies. There are dozens of other products out there that are expected to do the same, but probably are more user-friendly."

00:56:20] Is there a strong first-comer advantage in the medical imaging market?

RB [00:56:26]: Yes, that would be fair. I don't admire companies like Lunit in that, even though they're demonstrating great software, but it is such a vicious competitive market. Here in the US, for better or worse, there's a lot of consolidation, and that consolidation you see it in healthcare systems and radiology groups. A real-life scenario that's occurring all the time, we acquire centres and things of that nature. It happens literally every week. Let's say I'm a large radiology organisation, and we take over a group that has three locations and let's say they're using Lunit and I'm not. I'm going to use the opportunity to learn about the system. If I'm using a competitor, let's say Aidoc or some sort, and I look at it and go, "You know what, the system is good, but it's not better than my existing system, I'm going to integrate those new centres into my own system, and I'm going to get rid of Lunit because I'm not going to use similar systems at all." The threat is the consolidation. First in market and first in that I'm implementing, it's a good chance I'm going to keep it unless it becomes very cumbersome, too costly, or a product comes forward that is clearly a differentiator, meaning that Lunit is improving my accuracy by 30%, but then I hear from my peers that another product is improving accuracy after implemented by 50-60%, there's a good chance I'm going to start looking at them.

First in-house is a huge opportunity. When consolidation occurs, like I said, a provider could be Aidoc, let's say, being utilised at an organisation. If they are cannibalised by a larger group who doesn't use the system and they use a competitor, by no fault of that product, they're going to lose market share because they are going to consolidate. No major national company is going to try to reproduce an existing system. It doesn't make financial sense. They're going to standardise across their entire national network. An example of that is RadNet. RadNet has a subsidiary called DeepHealth. We mentioned it very early in the conversation. What they decided to do is not to go out to use a third party. They went out and basically built their own system. In some cases, they acquired existing companies and put everything under one umbrella. At the end of the day, they're going to use their system. Do you think they're going to use a third-party solution? No, they're going to use their own system. They basically cut out third parties 100%. Every time RadNet does a joint venture with a new healthcare system, if that healthcare system was using Lunit, well, guess what, probably Lunit is going to be out. If they were using Aidoc, their Aidoc is going to be out because they're going to use their own integrated system. That would be a perfect example of a threat to these third-party companies.

01:00:07] Have you heard about Lunit's second-read AI system? How do you gauge patients' reaction to it? Could you discuss its acceptance rate and other aspects?

RB [01:00:19]: Yes, I have. That's the part where patients have the opportunity to basically pay out of pocket for the system. I think it's high, especially in niche markets. Patients hear the coined term, "Hey, we use AI." They make the assumption, right or wrong, that it must be better. It's like that second pair of eyes looking at your study, especially if you are a high-risk individual, you want maybe that assurance. I think that is patients' right, you're going to see that. Organisations are starting to advertise that they use these systems. They want to let patients know, "Hey, we have great radiologists, but we're also using software systems that are AI-enabled that will enhance the radiologist's interpretation." In some cases, they just do it as an add-on value of it. In some cases, they charge the patient. It's an out-of-pocket expense and it doesn't go through insurance, but they say, "Hey, we did a mammogram on you, here's your report. Oh, you want us to use AI? We'll run it through our AI system, but we're going to charge you X dollars." That's a boutique service currently that you see many providers offer. Some providers just do it automatically as a quality control for their own group. Others, if AI is applied by the patient's request, they make them pay out of pocket.

ML [01:02:19]: Thank you very much, and thank you for your time today. Let me conclude this call with a huge appreciation to you.

RB [01:02:29]: Thank you for your time.

Transcription ends at 01:02:32 of the recorded material