EP17: How AI is Making a Difference in Breast Cancer Detection, with Tobias Rijken

Welcome to this new TechCaptains episode. So today we are happy to welcome Tobias Rageikon from the Netherlands and co-founder of Kheiron Medical since 2016, I believe, a company that raised $20 million to transform cancer diagnostics through deep learning. Hi, Tobias, how are you today? Great, thank you, Ron. Happy to be here. So, Tobias, I want to start really. We met each other about 10 years ago in Hackathon, in a train that was going from London to Edinburgh and back, and you hadn't started a company at the time. I think you were studying machine learning at UCL. That's right. So what's happened since then? What's happened since then? So, yeah, I think... We, as you said, we met on this hackathon on a train, which I thought was a great idea, actually. I think I was still in my figuring out what I want to do with machine learning. And I was, at the time I was at UCL, the program was called Computational Statistics and Machine Learning, which really was a fascinating program, and especially around that time, because just to put this into context, this was around 2014, 2015, I believe, right Ron? And this was just the beginning of a new, the deep learning revolution had just kicked off. So, right. AI goes into these, goes through these cycles of AI winters and AI summers. And this was the beginning of a new AI era. DeepMind had just been acquired. And DeepMind was founded at the group where I was at, the Gatsby unit and the CSML group at UCL. And what was really cool was that a lot of the researchers at DeepMind still had academic positions at UCL. So there was a lot of cross-pollination. My supervisor, for example, was one of the lead scientists on AlphaGo. David Silver was still teaching the reinforcement learning course at UCL. So that was really fascinating. And one of the things that I enjoyed about machining in general is the ability to solve real world problems. At that time, I was really looking for, okay, what is a big real world problem that I can address with machine learning? My attention got on reinforcement learning. I worked with my supervisor on building a system to learn how to optimize traffic controls. So imagine you have a reinforcement learning agent that every junction in the city cannot learn how to optimize. optimize traffic control and flow and minimize waiting times and CO2 emissions. And it worked surprisingly well. And that's when I thought, okay, what's next? So was it your idea? Was it your idea to start? Or how did you get that idea? I guess moving from Amstam to London was quite a... shock in terms of my day-to-day interaction with traffic. I mean, in Amstam, you can cycle everywhere within 20 minutes, which is great. And London is very different. So I thought, okay, and this was also the time when the, what's this called, the low emission zones were introduced. So I thought, okay, is there a way that AI can help optimize traffic to minimize waiting times, but also minimize CO2 emissions? And well. I found a simulator from a transport research group and decided to build an agent that can learn to do that. So it was a fun project, but I realized that fundamentally the problem in traffic is if you just have too many cars on the roads, then your AI can be as good as you want, but the system will get congested. And so I think there's a more fundamental problem. in traffic rather than optimizing the control flow itself. So, I mean, I grew up in a very medical family in Amsterdam. And I think this was a time when I think healthcare was starting to open up, or at least what can we do with the data available. So I joined a company at the time, which was still quite a small company, it was called Stratified Medical. It's now known as Benevolent AI. So Benevolent, I was one of the first six people in the machine learning team. And what Benevolent was doing was, or still is doing, using AI and machine learning to help drug researchers find new drug targets. And now they're actually having their own drug pipeline and bringing drugs to market. But the idea was, what I was working on was from academic literature can we find some sort of, can we extract knowledge from that literature and help drug researchers with clever tools to find drug targets more easily? And I thought that was fascinating. What I wanted to work on, however, was something a lot earlier in the cancer pathway or in any disease pathway in general, because One of the big problems in healthcare these days is that we're spending a lot of money on therapeutics, treatment and drugs, and not enough on the earlier stage of the disease pathway, which is early detection and prevention. Because if you find disease early, the cost of treatment is orders of magnitude lower. And this is when I decided to join Entrepreneur First and build a company in that space. So how did you find the experience at Entrepreneur First? Yeah, I mean, this is quite early in Entrepreneur First journey. We were cohort number six. So EF itself was only three, four years old. I think Entrepreneur First is also still figuring out what it was going to be. The whole idea of putting 100 technical people together in a room and that they will find their co-founder build a very impactful business at the time seemed quite nuts. But I think it's also one of the most important aspects of starting a company is who is your co-founder going to be. And that's a very hard question at any time to answer because finding a co-founder is... It's hard. I mean, you might know the perfect person. It's like a marriage. It is. It's like a marriage, isn't it? Yeah. And, and, and most often, uh, co-founders are not readily available, right? You might know someone who you think is great, but they're in a job and maybe it's not the right time for them to start. And what EF does is it creates this pressure cooker where at least you have a hundred ambitious technical people who've already made the decision. They want to start a company right now. And so constrain that pool and then work from there. Uh, so I met, that's where I met Peter, my co-founder. And I think we very quickly, uh, bonded over, uh, sort of a shared ambition. And, and, uh, both of us come from very medical families, uh, but already sort of odd ones out in the family doing math, computer science. Uh, Peter had just finished his PhD at Oxford and high performance computing. Uh, but his mother is a radiologist. Uh, and so he had spent a lot of time. Okay. Yeah, he basically always joke that Peter learned how to use a CT machine before he learned how to use a computer, which is actually not a joke. It's true. It's true. The entrepreneurs first then is, is this the first I'd ever heard of this organization when I saw your profile? What are they doing? Are they just matching people with technical skills? Is that what they focus on? Yeah, I mean, it's, I would describe it as a It's not really an incubator or an accelerator, it's sort of a company builder. And they take cohorts of individuals and they essentially invest in technical talent, pre-idea, but also pre-team. So the goal of Entrepreneur First is to match really strong technical people who want to build a company together. And they take a couple of cohorts. It started in London a couple of years before we started the company. But now they already have cohorts in Paris as well and Singapore. And it's grown to be quite a big organization. But I think now you're starting to see the successes because some companies from the earlier cohorts are now starting to mature. So in the UK, Tractable is one of the bigger names that came out of Entrepreneur First. our cohort, there's also Accurex doing a lot of work in the NHS. And so it's now, I think we're now starting to see the success that the model has. But of course, company building takes a long time, decades or more. So I think the true proof is still out there. But yeah, it's essentially get people together and help them form a team and build a business. And can you work... Can you walk us through maybe through the business model of care and medical, like who are your clients? How do you make money out of cancer diagnostics? We are essentially addressing the cancer pathway. And what we want to do is we want to fix cancer with a better AI driven diagnostics. So one of the big problems that we are seeing is that cancer care is actually bottlenecked by diagnostics. It's an information problem at every stage in the cancer pathway, all the way from detection to diagnosis. treatment planning and treatment, there's what we call an information problem. Either information is inaccurate or there's missing information or the information is ambiguous. And that leads to suboptimal decisions of cancer care. This is something that can be solved with AI. So what we are doing is building AI along the cancer pathway to improve the quality of the information at that specific point, whether that's a better diagnosis or more accurate detection. And we sort of longitudinally model how cancer progresses through that pathway so that we can improve cancer care. I mean, that's quite a big vision. But what we did is we started all the way at the beginning of the pathway, which is, before you can do anything else, you need to actually detect is there cancer here? Yes or no. So detection is step number one. And one of the hardest breast cancer detection tasks, but also one of the most established cancer detection tasks is breast cancer detection. Many Western countries have an established breast cancer screening program. So essentially the task that radiologists need to perform is a mammogram is taken of women within the age group 55 to 70. It sort of depends per country what the window is, but they get every year, two years or three years, they get a screening mammogram. And the radiologists need to look at it and basically determine, um, should I call this woman back for further examination? Yes or no. It's a binary task. Um, the problem, however, is that the, there's a huge shortage of radiologists at the moment, um, and in the, in the UK, it's even coming to a stage where, uh, if we don't solve this crisis, uh, some some screening programs will no longer be able to continue. The Royal College of Radiologists has been speaking about a workforce crisis for a number of years now. We simply don't have enough radiologists to do this work. And the way cancer screening works, the gold standard is you always have two radiologists looking at every single case because it's a very hard task. It's very well-defined. but still a very hard task. So two radiologists look at this independently. When they agree with each other, great, you do whatever they agree on. When they disagree, a third radiologist comes in to arbitrate the case. Now the interesting thing is this is a screening distribution, which means that roughly 1% of women has cancer, and the vast majority of cases are normal. So in most of the cases, the two radiologists actually agree with each other because those are the easy normal cases, but you still need two radiologists. So what we have developed is MIA, the mammography intelligence assessment, and MIA can essentially perform the task of one of those radiologists. So does exactly the same task, mammogram comes in, is there a cancer here, yes or no. And then the workflow doesn't change because now you have the radiologist and MIA, if they agree with each other, great, you do whatever they agree on. When they disagree, you still get the arbitrating radiologist. So there's always a human in the loop, but now we can... automate all the reads of the second reader. There will be a bit more arbitration, so overall you save roughly 45% of reads. But this is high volume cases. In the UK, we're screening more than two million women a year, and we can automate half of the necessary reads here. So that's a bit of a context. This work, we've now started to see real world evidence that we This was also featured on an article in the New York Times earlier this year on some of the work we've done in Hungary. And now we're rolling this out in the UK. We're currently rolling out in 15 UK screening units, which is in terms of volume is roughly 25% of the UK screening volume. I was going to ask a question which you kind of answered, which is, you know, why did you focus on... on breast cancer. I think you kind of answer that by saying, you know, there's established programs, there's a set of data with outcomes that you can train models on. So what do you think you'll do next? I mean, it sounds like you're getting to the place of cracking this problem. What do you think you would focus on next? So yeah, great question. So as I mentioned, detection is really at the start of the pathway and it's only one cancer. The next step for us will be go to find more cancers. detecting more cancers in different modalities. So breast cancer is done with mammography, a type of x-ray. The next step would be finding cancers in CT scans, which is basically, yeah, it's basically everything, all the cancers that occur from the neck to the pelvis, you can detect in CT scans. Okay. And then you have this other product, so I was just looking at your various products. So you have this thing called, is it mere IQ? And I watched a video last night and I was reading about that and it said that, you know, 47% of technical recalls are due to incorrect positioning of the breast, right? This is put on a plate, I think, isn't it? But what I was trying to get was how does your service work then? Is that a real-time analysis that you do of a scan that's taken? How does it sort of fit into the workflow and the experience that women have with using it? In short, it depends a bit on the value proposition. So how we are thinking about this is we are building a suite of different solutions along the pathway, sort of longitudinally. We don't believe in just building a single point solution because point solutions only offer a subset of the value that we can deliver, but also it's easier to be replaced, right? We believe that the true value comes from having multiple solutions along the pathway. Take for example, image positioning and Mia. There's no use sending a case to Mia to be asked for an opinion if it's poorly positioned. So the two together will create more value because they basically reinforce each other. So on what the experience is for the women, it's mostly running in the background. So most women will not interact with Mia directly. we work is we need to integrate with hospital IT systems in the hospital. And of course, data privacy is of the utmost importance here. So what we do is we have a gateway that we install on-premise in the hospital that essentially integrates with the system where the images are stored and the systems where clinical information is stored. We de-identify that data and then we send it to our cloud. And in the cloud is where all the AI algorithms run. That's important for us because that makes it a lot more scalable, especially because inference of AI models are quite resource intense. So we want to be able to control that and use it as efficiently as possible, those compute resources. And then we send the result back and in the gateway, the result then get recombined with the patient data and get sent back to the... to the hospital IT systems. So your cloud servers doesn't know who is the scan for. It just has an anonymous ID. No, exactly. Yeah. And so speaking of, can you maybe walk us through a bit through the, so you say, you spoke a bit about the technical architecture. But if you can go deeper into the language, the frameworks, how do you handle the infrastructure, like your your AWS cost if any or anything like that. Yeah, no, I'm happy to. So maybe just as a bit of context, I think the ML stack is quite new and very quickly evolving. So for example, when we started the company, most of the ML... frameworks did not yet exist. So TensorFlow, for example, didn't exist when we started the business. I built the first models in Lua Torch, which was one of the frameworks being used at the time, which was an absolute pain, but luckily TensorFlow was released not too shortly afterwards. So right now we're primarily sort of Python, TensorFlow based. I think maybe more than... than other companies, we invested quite heavily in hardware. So of course, training AI models requires quite a bit of compute power. And we basically realized quite early on that doing that in the cloud would be prohibitively expensive. Even with all the free credits you can get as a startup, we burned through $100,000 worth of credits in a period of three months by using, yeah, like eight. AGPU clusters in the cloud cost you a lot. I can tell you that. So what we did was we invested quite early on in building our own infrastructure, our own GPU cluster. We built initially in the coworking space that we were in. Now it's in a colocation space. And I think that's probably one of the best financial decisions we've made. It was quite a hard discussion we had to have with our seed investors, because basically of the fundraising round that we raised, we, we told them we need to spend $125,000 on a gold, golden computer because the NVIDIA DGX was gold plated back in the day, but that computer is still running today. Six years on, we're still training a lot of our models on that machine. which is only a fraction of the cost that would have taken us to do this on AWS. So yeah. That's really interesting. Yeah. So you, so just to be clear, you basically built your own server with GPUs in it that you purchased and you've put it in, you've co-located that in a data center somewhere. And the GPUs, the cards that are in there, you've not changed them. You still have the same Nvidia cards in there that you put in six years ago. Interesting. Because there's this idea, right, isn't there, that the GPUs now, I mean, Nvidia stock price has gone crazy because of CHAP GPT. There's this idea that you continually need the latest and greatest card to be at the forefront. But what you're saying is that, no, if you have something purpose-built and you know how to use it, that you can actually sit on the same hardware for quite a bit longer in your use case. In our use case, however, I would say we are stretching the limits of the cards quite significantly. Most importantly, GPU memory. The cards that we have, the largest card of 32 gigs of GPU memory, I think the latest H100 cards have 80. That does allow us to train larger models, et cetera. So we do need to think of how we refresh and how we do the infrastructure upgrade. But I mean, by all means, this was the way more cost-effective way to get GPU resources. It would have been... impossible to do this with a startup budget, it would have been impossible to do this on the cloud. And is there anything else that's special about these servers other than the fact that you kind of built it yourself and chose those video cards to go in there? Is there anything else? Because Google put a big thing, didn't they, about their own servers. It's kind of an interesting, you can see some interesting articles about their own servers that they built a long time ago, that are very, very custom to them. How custom are your servers? So they even went one step further. They've built their own chips. the TPUs. So we're not going that far, but we have designed to service ourselves. I think one interesting question that we have been thinking about over the years is how do we combine the benefits of having your own on-prem compute with the scalability of the cloud? And so we've always operated in a bit of a hybrid model, but that creates a lot of challenges. So because the data sets that we're working with are very large and the data primarily resides in the cloud, but the compute happens down at the data center. So thinking about how we cache our data sets, how do you can have a fast server, but if you can't load the data onto the server fast enough, then you're bottlenecked by bandwidth and IO. And so what we've done now is most of our data still resides in the cloud, but we've built a huge. SSD cache to make sure we can load data onto the machine fast enough. So it's created its own set of challenges, but again, I think it's, it was, it's worth the investment for us. So that's, this is primarily what we do internally. The product stack is, we try to keep as simple as possible. So it's Python and TensorFlow on the backend. And then the gateway that's deployed in the hospitals is written in Kotlin. And that's what we've been running for the last couple of years. Tobias, like we have a small tradition in the podcast that, you know, the guest introduces a book they really liked. And the one you mentioned you wanted to present is called Factfulness. 10 reasons we wrong about the world and why things are better than you think from Hans Rosling. So tell us more about this book. Yeah, it's, I love the book. It's actually quite a while ago that I read it. So Hans Rosling was a famous professor in public health. Unfortunately, he passed away a number of years ago. But he also became quite famous for his TED talks where he used very intuitive visualizations to help people understand data and facts. What I like about the book, as the title also suggests, it's quite positive. He shows us how the world is actually a lot better than you often think if you only read the media and the news. But I think what was quite an eye-opener for me is he helps us understand why we're often wrong about certain facts because of certain intuitions or beliefs that we have, or because of incentives that we have or other people have because of their job or the work that they're in. And so he breaks that down in quite an intuitive way of what are all the biases that sort of exists when we look at these. these facts and just help us understand why we might be wrong about certain things. It's a great read, highly recommended. Very interesting. And Gareth, you had another question, I believe. Yeah, so I was reading about, I think, the pilot that you have run in Grampian, and there was a quote from a doctor on there who said, it's great at detecting breast cancer, but unable to review the findings in the context of past scans and medical history. So I'm guessing at some point of those two, are you gonna be addressing the issue of being able to bring in data from past scans? Is that something that you're looking at doing? Yeah, absolutely. We're already doing that in our R&D model. So our models can look at, this is called the prior and taking the information from the prior into account. So from a machine learning point of view, we have a solution. I think here it's often the practical solution that, or the practical questions that we need to consider as well. Often the prior is not available because maybe the woman in question moved between two different regions of the country and the prior is not available. Or what do you do when there's been a database migration and all the old scans now reside somewhere else? So there's a lot of practical considerations. that we're looking at, but, yeah, priors, it's very, it's the natural step to take into account which we're already doing in an R&D point of view. Because radiologists look at that as well. If you look at a radiologist, they always compare what am I looking at now, what was the priors scan looking like, especially because, well, if there is fast growth, looking at the prior will help you detect that very quickly. I mean, the NHS, obviously I've been to NHS doctors and I've had x-rays and things like that. And then I've been to private doctors here. And I remember there was one I went to, I had a weird chest problem, so I had a chest scan and he emailed me the actual x-rays afterwards. So I actually received them and I was thinking, you know, why don't, why isn't that normal practice? You know, why don't I just own all my images and I can just have them, I can give them to whoever I want to give them to, right? What's stopping that happening in the NHS, I wonder. Yeah, it's a great question. I can talk about this at length, but it's funny because I think that the model that you point out is sort of a more decentralized way of keeping healthcare data, right? Our healthcare system is decentralized by default because we don't have advanced IT systems. So actually the patients keep all the records on them. And when they go to a doctor, they bring their whole history with them in a file and show them, Hey, this is all my data, which, which I thought was fascinating because in the U S in the, in the UK, it's so hard to get access to your own data. Yeah. We once supported, um, one of our, one of our advisors who was on the going cancer treatment and, um, and asked for if we could take a look at his scans. And he asked his NHS, various NHS trusts to send him his data. And essentially what happened was I got a stack of DVDs sent to my house, which I had to upload. Well, first of all, my Mac does not have a DVD drive. So I had to buy an external DVD drive. Then they were encrypted. Which, hey, great, I understand that you want to encrypt data, healthcare data when you move around, but the encryption software could only be run on Windows. So I couldn't unencrypt the data on my computer. I send it back to a cloud folder and then the advisor could do that himself. But I think the frustrating part as a techie is that technologically, this is a solved problem. Right. But. in reality in healthcare systems, it's not. So I think there's still a long way to go there. But I think you're going to be hugely disruptive with this because it's slightly scary to hear you talking about the shortages that are potentially upon us in the NHS. And I've read things about that. But interestingly, I worked in the US for quite a few years and I worked for a medical systems company based in California. And at one point we went to do some work in Hawaii. which was with some radiologists that were based in Hawaii. I didn't know anything about that world at the time. And I flew out there and they told us, oh, what we do is they only work three days a week. And what they did was they were reading scans from the East Coast at night. So when the hospitals were scanning patients, they were, it was the middle of the day for them, or earlier in the day for them, later in the day, I guess, and they were actually reading those scans in Hawaii and then sending the results back to these hospitals on the East Coast, which was an expensive thing for these hospitals, right? So. I think your tech is going to be massively disruptive. It's very exciting. Yeah. Especially if you think of the parts of the world where people don't have access to adequate healthcare at all. I mean, breast cancer screening is very well established in the Western world, but there are many countries that don't have a screening program at all. And what that means is that by the time the woman shows up to the hospital, the breast cancer is already in a stage three or four. and survival chances are much lower, cost of treatment is eight times higher. So finding cancers early matters, but without radiologists, you can't. And that's, I think, where a huge opportunity for this kind of technology lies. Very interesting. I mean, I learned a lot. I think what you're doing is really fascinating, Tobias. So on this world, I think we will close this post-gaz for today. So thank you so much for spending the time with us today. Thank you. Thanks for inviting me and great conversation. Yeah, thanks, Tobias. Yeah, thank you, everyone. And I'll see you in two weeks.