Co-Founder of Annalise.ai Aengus Tran on Using AI as a Spell Check for Health Checks - Ep. 207

0:10Hello, and welcome to the NVIDIA AI podcast. I'm your host, Noah Kravitz. My guest today is Dr. Angus Tran. Angus is the co-founder and CEO of Harrison AI. Harrison AI is a clinician-led healthcare AI company using deep learning to tackle one of our biggest global challenges, the inequality and capacity of the healthcare system. One of Harrison's ventures is Annalise AI, which grew out of a partnership born in early 2020 with iMed Radiology to develop comprehensive solutions across radiology modalities. Dr. Tran is a world-renowned expert in deep learning and healthcare who was originally trained as a doctor at New South Wales University in Sydney.

0:50Dr. Tran, thank you so much and welcome for joining the NVIDIA AI podcast. Thank you for having me. Pleasure to be here. So why don't we start, if you would, tell us a little bit about your own background and how you got involved in the field of AI and obviously as pertains to healthcare. So I came to Australia about 12 years ago as an international student. And growing up in Vietnam is where I come from, Ho Chi Minh City. My father was a math teacher and he was quite innovative for his time. So he introduced programming as a subject into high school in Vietnam. Back in the day, it was no computer science as part of the curriculum.

1:31He introduced Pascal and was teaching, was coercing Math Olympiad kids to go over to the dark side and do programs. Growing up, I always was exposed to algorithm developments and programming from a very young age. And I always thought I would grow up to be a computer science engineer. But when I went over to Australia and studied, I really fell in love into medicine because one of the things I realized is computer science was a tool that can be used to solve really complex problems. And what other more worthwhile problems than to look after your fellow human being and medicine? And unfortunately, unlike programming where you can learn online and self-taught, medicine is really a human endeavor, right?

2:17It's an experience. And you couldn't read that from a book, unfortunately. So I ended up in medical training at University of New South Wales, Spent six years there studying the art and craft and the practice of medicine and really just fall in love with how impactful medicine can be for a community and for the individual. During this whole time, I was still very much obsessed with programming. I tried to start a failed startup from the dorm room like you would. What was the focus of the failed startup? The startup was in the early days of ticketing before phone ticketing was cool. So it was like a code-based event ticketing solution with my friends.

2:57I also dappled in programming while I was in my medical training on the side. And this is around the time where a lot of the breakthrough in machine learning and deep learning come through. So you would hear things like the Offergo program. You would hear about things like Jeopardy, the AI. So when I was growing up, I was programming chess engine and other algorithms. And when I saw this breakthrough, I said, something was different here. Because when I was programming as a kid, this was a very, very hard problem. You was doing like min-max search. And it didn't do these things. So clearly, something was very different.

3:35So that's when I said, I have to take a deeper look at this. So when I went down a rabbit hole in machine learning and deep learning, And I take a lot and use my saving at the time, very little saving, be that. Bought some NVIDIA GPU, actually. Gave myself a deep learning tower. Borrowed some money from my brother, Dimitri, at the time, built a bigger tower. It was 1080 TI at the time. It was stay-of-the-art at the time. I get really deep on that, just start competing on Kaggle. You know, I did a lot of competitions there. Did quite well in some of them. And yeah, I mean, and just start being fascinated about this notions of applying AI in healthcare.

4:15So that's how I got started in the field. Around the fifth year of my training, in my final training, I got involved in a project that people commonly know Harrison and myself as is the IVF AI technology. It made a lot of news at the time because it was technology that can look at human embryos, the video of it, and decide if the embryo is likely to create a baby or not. Right. So IVF, in vitro fertilization. Yeah. Yeah. In vitro fertilization. So there would be 20 embryos. There's videos of it. And I develop an AI system that can classify them into pregnancy or no pregnancy. And that rapidly get adopted and deployed across Australia and then the world, right from where I was at the time as a medical student.

5:03and really feeling that medical impacts, right? So I looked at that story and I said, wow, if I was practicing as a clinician, I'd be lucky to see like 100 patients and really make an impact in their life over the course of a year. And here, you can touch thousands of patients by developing AI tools that scale the knowledge and expertise of medicine. So I thought, this is a really different way of practicing medicine. And that's how I ended up roping my brother, Dimitri, to start. You know, he quit his job at the time as head of innovation from Ramsey Healthcare, a big private health company here in Australia.

5:43And together we started Harrison with the mission to scale global capacity of healthcare with a suite of autonomous AI system in multiple medical domains, starting with radiology. And then now we're in pathology and other things. But, you know, radiology has been our focus for the last few years. Right. So I think we need to shout out Dimitri because he lent you the money to build that. It was the second tower, right? To build the bigger tower. And then he pulled him away from his big time job at the big healthcare company to start Harrison with you. That's fantastic. So you mentioned the focus on radiology now, and you've got a product, or I guess it's sort of a subcompany that I mentioned in the intro, Annalise AI.

6:23What does Annalise do? What's its mission? So Annalise was our first attempt in autonomous AI system in diagnostic medicine. And the front door of healthcare is radiology. Most of us, across our life, if we ever needed to engage in health system, we most likely have some sort of imaging. I had an x-ray done a week and a half ago as we recorded this, in fact. So I can vouch for that. Yeah, no, I mean, and that experience spanned a billion times a year across the world. Yeah. So Annalise's mission is to scale the global capacity of radiology diagnostic with autonomous radiology system. Okay. So the reason why that mission is so worthwhile is because one of the biggest shortage in medicine is radiologists.

7:15For example, in Australia, where I come from now, we have 2 ,000 radiologists looking after a population of 20 million people. Wow. Why the shortage? Well, the reality is because it takes 12 years to train radiologists. So if the audience plays StarCraft or other like Warcraft, when you press build and a unit gets built and then a marine comes out and then there's a progress bar, that progress bar is 12 years for radiologists. Right, right. Right. So for me, it takes six years to train as a doctor, and another six years of subspecialties to become a radiologist, which means that if the government put all the money in and say, radiology training is free now, anyone can get through, we get the result of that 12 years from now.

8:02So it's a huge gap. There's a gap, yeah. Yeah, it's a huge gap. And it's also something that's incredibly manual, right? So most people don't realize this, but when you have your X-ray on, That image is displayed on the computer monitors of some radiologists somewhere, and they themselves look at it, analyze it, and decide what goes into those reports, right? So it's a very manual process, not at all automated, because of that we have a huge shortage. If you think about Australia as a shortage, our neighbor, Indonesia, just across the ocean, they also have 2 ,000 radiologists for a population of 200 million.

8:45So if aliens lands on Earth, the thing that would say is you guys don't have enough radiologists. That would be the acute observation that it made. and for analysts, you know, being able to build this AI system that look at these images and produce diagnosis such that, you know, we can speed up radiologist time, but also allow clinicians to make better and faster diagnosis, you know, when there's no radiology coverage such as at either in rural region, what an incredible impact that would be for healthcare. So how does it work? Yeah, if you have an X-ray, many of the audience can even try it themselves.

9:22We released this as a demo on our website. Oh, no kidding. Yeah, it's analytics.ai. I don't mean to make this about me, but I haven't actually gotten my report back yet because I had the x-ray on a Friday. My physician was out for a week. So this is great. I'm going to try this when we're done. So forgive me for interrupting. Go ahead. No, it's a very common experience for many of us, right? In the UK, for example, the wait time for diagnosis is about two weeks. They met Somebody, you know, worry having, you know, a chest x-ray for pneumonia, for cancer, for a diagnosis. Imagine the anxious way they go through.

10:02So the technology is that, like I said, the demo is on analyst.ai slash web demo. You can even upload images and see what it says. It's not a diagnosis, it's a demo. But this is the same technology that clinicians all around Australia and the world are using today. You take a chest x-ray or CT brain, and we have a comprehensive AI system that look at those images and produce a suite of 124 to 130 different diagnosis. And these are the same 124, 130 features that a radiologist would be looking for. Now, the difference is because it's an AI system scalable by GPUs and clusters of it, right? It can run at massive scale.

10:42So for example, instantly, a couple of seconds after the image is taken, it can look at it. And if there's a critical findings, it would flag that on a wait list. Imagine someone going through your email and pick out the most important email and flag it. That's the same sort of experience for the radiologist. But the other thing also is it acts as a safety net. Like I said, this is very human and manual process, meaning error is very possible. There's a study that said lung cancer is missed in something like 5 % of all x-ray. Wow. So it's a huge amount of errors. So the AI system acts as spell checkers, if you like, for radiology.

11:21Well said. They make their diagnosis, and the AI would go in and make sure that nothing important is missed. And that works across the chest x-ray and CT brain, which is two of the most common medical imaging modality in the world right now. Right. And so the idea, which is, if I have it right, which is a common theme amongst all of the, you know, I say successful, whatever the measure is, but the successful implementations of AI is you have a human in the loop. It's not replacing radiologists, but it's enhancing their ability to, you know, to make more correct diagnoses basically, right? And part of that is that safety net and part of it is flagging those critical findings, as you said.

12:00What were some of the challenges you faced or are still working on if the case may be in developing, Annalise? Yeah, so for these tools, one of the challenges for us is that the minimum algorithmic performance is very high. So if you were building, for instance, a traffic-like classifier, if you were building a hot dog or not classifier, you may get away with a performance of 80 % to 90 % because the consequences of an error is the fear. Whereas in medical diagnosis, this AI system has a very high bar of accuracy because ultimately you are helping people diagnose for fractures in pneumonia and cancer.

12:47So we spend a huge amount of time trying to solve the accuracy issues. And that comes from a number of places, right? So number one is the data set, creating, curating a massive millions of images data to train our AI system. Unfortunately, there's no ImageNet for ChessX, right? These data set, we need to build and grow ourselves. We also have a huge data annotation efforts at Analyze. So unfortunately, we can't get anybody off the street to label for the streetlight or the hot dog. We have an army of radiologists to do this. So Analyze engaged with hundreds of radiologists to go through these images and very meticulously annotate for the present of these findings in very high accuracy.

13:32Many times these images are annotate three times over, five times over, such that the consensus can be used to train the AI system. So we are squeezing the 0.2, 0.5 % of the accuracy out because that when you're multiplying to a billion chest x-ray a year, it translates to thousands or 10 ,000 of medical diagnosis and better outcome for patients. So that is one of the key challenges that Annalise worked through, which is the development of this highly accurate and comprehensive AI system. I'm speaking with Dr. Angus Tran. Angus is the co-founder and CEO of Harrison AI, a clinician-led healthcare company that's using AI, using deep learning to improve the access to healthcare and quality of healthcare globally.

14:19And we're speaking specifically right now about one of their sub-ventures, Annalise AI, which is focused on radiology. As we've been discussing, is using deep learning techniques to help radiologists make more and more accurate diagnoses across some, I mean, literally life or death situations, as we've been talking about. We mentioned kind of at the top that Harrison started out with the IVF project and has moved on and has focused a lot on radiology right now. Now, what are some of the other things, aside from Annalise, that either you're working on or might be in the plans kind of going forward?

14:57Because it seems like from the company's background and certainly your background and looking at the website that ideally your reach will be broader. I don't want to say than just radiology, because it's obviously a huge need, as we've been talking about. But what else is Harrison working on? Yeah, so our mission is to scale healthcare with autonomous AI system and enhance the clinician's ability, right? Within that, you have radiology. And radiology, as you correctly identified, is a very deep field. We have much more work to do to solve a bigger part of radiology. we start with the two biggest modalities, including chest x-ray and CT brain.

15:37We are working on other modality like CT chest, breast imaging, imaging of the joint and musculoskeletal system, imaging of the abdomen and the pelvis. So there's a lot more work that we still aspire to do and to drive further automation in radiology. The other thing that we recently started is in the field of histopathology diagnosis. So we have a sibling company for analysts, if you like, called Franklin AI. Franklin is new, is still in product development at the moment, but we're looking to share the same vision, but in the field of histopathology diagnosis. And what is histopathology? So histopathology is when a clinician make a biopsy and get a sample of a lesion, potentially a cancerous lesion.

16:24They then cut very thin slices of it, like a salami slice, and then put that under a light microscope to inspect for the presence of cancerous cells. So we are building an AI system that can help the pathologist in analyzing those gigapixel images to understand the presence as well of cancer, but also the outcome prediction for the patient. How similar or different is Franklin to Annalise in terms of, you know, the technology and your development process and the milestones and hurdles that you're facing? You know, at the core of it, there's some very similar patterns, which include building a high-performance computer vision system, large-scale data set management and annotations, but also some of the rigors around quality and accuracy as we previously discussed.

17:15So actually, Harrison developed a lot of shared toolings, a lot of shared technology that get used across products, but also across entities like Franklin and Analyst. So Harrison developed a suite of libraries. We call Harrison AI Machine Learning Suite that include very common building blocks across this. So we've done many different projects now solving diagnostic AI system automation. and we look to be able to transfer some of that technology across projects and hopefully to accelerate that development over time. There's a lot of talk, I was going to say in the States, and I can extend that at least to some parts of the UK because I've been talking to some folks in the UK recently more than usual.

18:01But there's a lot of talk certainly in the US about AI and about data privacy and ethical implications and, you know, kind of a mix of some outright fear about AI systems and what's happening. But I think more of a focus on seeing the good that AI-enabled tools can bring to the world and then trying to figure out how does the data fit it. And so in the case of, you know, generative AI has been getting a lot of press, at least over here, I think globally, but you can correct me if I'm wrong. And so there's a lot of talk about, you know, Companies scraping the web and using that data for their training sets and then the ethical implications for artists and writers, for one.

18:47What are some of the ethical issues that you're facing that Harrison is grappling with in the healthcare industry more broadly when it comes to using AI in healthcare? I think for data set that we use, they all de-identify an anonymous data set. And there's been a lot of work being done to ensure that this meets all the requirements of the law and being compliant with those regulations in data privacy and health. I think, you know, for me, thinking about the ethical challenges of using AI system in health is actually the converse, which is, is it ethical to not use AI for medical diagnosis? So I imagine like a thought experiment for you.

19:34I imagine you were on a plane about to take off or Sydney or Australia, maybe you come and visit us. Sure. And as the plane rolled down the runway, the intercom come up and say, dear passenger, the autopilot system is down today and I'm feeling adventurous, so I will be taking off manually today. Enjoy the ride. How would you feel about that? As soon as the pilot said they were feeling adventurous, I'd be trying to break my window out and jump outside of the plane. Right. But what if this is not with no disrespect to the pilots out there? I just get a little I get a little worked up quickly. Yeah, that's essentially what we're doing right now in medical diagnosis.

20:12Right. We're saying, you know, there's AI machine learning system that's capable looking at chest X-ray and CT brain. And, you know, we have publications in, you know, credible medical journal like Lancet that shows that the system simply is more accurate. human using those AI systems simply pick up more misdiagnosis, pick up more cancer and other conditions. So, you know, one of the things I'm grappling with is, you know, if I have a chest x-ray done today and if my family members have a chest x-ray done today, is it ethical to have someone looking at that without the protection and support of that?

20:49It's the same as, you know, is it responsible to write a script for a podcast without running it through a spell checker? You You know, that's what I'm thinking about. So suddenly there's a deep sort of utility view in AI system for medical diagnosis. And of course there's concern. And we as a society just really come together and have those really deep debates about that. I would say, you know, other industry using AI system should learn from the medical industry because AI in healthcare is actually very well regulated and there's very robust discussions on what safe usage looks like. How should we validate it?

21:28How should we get regulatory clearance for it? How do performance get measured and benchmarked? Because of the rigors of healthcare, as I said, the accuracy requirement is very high and is very high stake. This actually wrote the playbook for what good AI implementation and mission-critical application look like. So I think other areas like chatbots and other systems, you know, we don't need to reinvent the wheel here. We need to look at medicine. Look at the medical playbook. Like what are we dealing with? Right, right. Be it, some of it is quite overdone at the moment, right? A lot of it is very high bar, but it's good sort of rule books on what safe and ethical AI usage look like.

22:08What do you think is next? And this is a broad question, so, you know, tackle it from whatever angle makes sense to you. But the role of AI in everything, I think that folks who have been listeners of this podcast would know, but people have been working with AI in some capacity. The past year and a half hasn't been this brand new thing as it's been sort of in the mainstream, the explosion of chatbots and, you know, things that users could interface with language models really opened a lot of people up to this idea of, oh, AI is here. But people have been doing things with AI for years and years.

22:44The healthcare industry, as you said, certainly among the leaders of using the technology. Where do you see AI and healthcare going next? Is it, obviously you mentioned Harrison, you're focused on scaling, right? So you're providing these services and helping healthcare practitioners be able to do more and in turn better healthcare for more people. And it's a matter of scaling that out. And we talked about diagnostics. As an industry at large, is this kind of the trend that you think is gonna happen? It's a matter of, you know, doing these things a company that's like yours are already doing and refining them and scaling them?

23:21Is there some new, you know, thing coming, whether it's generative AI or some other AI-powered thing that, you know, the first thing that always comes to my mind is, you know, increased individual personalized X, right? So in the case of healthcare, maybe it's better preventative care because my phone can use AI to track myself, you know, my vitals and everything and give me better preventative care ideas. What are you thinking about? I actually think about it in sort of two buckets, right? So one is what I call AI inside, and then the other bucket is what I call AI first. So AI inside is what we're working on now, which is we are doing the tasks that previously done by a human.

24:08We're just doing that. We just fix some of the bugs, right? Some of the bugs is the number one is the scalability, as we touch on. The fact that it required 12 years to build a radiologist means that the scale is limited. The other buck is actually the fact that people go to sleep at night and they go play golf in the weekend, meaning that the availability of this service follow human availability. And unfortunately, people don't get sick on a circadian rhythm. I mean, you're equally likely to have a car accident requiring a CT head at night. and you're equally likely to get kicked by a horse in a farm in rural region versus falling down a stair in the middle city.

24:51So healthcare and medical needs is universally needed, but the availability is following like some of the geographic and circadian distribution, right? So the other buck that we look to fix in the AI inside on just tasks that people are doing now is the accuracy, right? The variability. So clinicians wake up in the morning they got their morning coffee and they're top tip top performing they're unlikely to make a mistake right but they have a burrito at lunch a very hairy burrito and you know melatonin and stuff kicked in they're going to miss more things so the lesson here do not go and have x-ray in the middle of the day in the middle of the day and then you know during the day not where there's good Mexican food yeah yeah there's variability in diagnosis right so it's a non-determinant system.

25:42So there's a huge opportunity in just fixing some of the bugs with current system and scaling it. The other area that you allude to with AI first is new knowledge, right? Like discovery. And I think that's equally exciting. For example, you know, there's pan in our data today that, you know, we just simply haven't seen yet. For example, you know, can we predict the present of cardiac risk, you know, risk of heart attack from looking at, like you said, you know, personalized companion device data or looking at unexpected modality like the retina to predict risk for cancer and other conditions. So I think there's going to be a huge amount of discovery coming through on that side of the fence as well on what are new and different ways of doing medicine that's only possible with AI.

26:32So I think both need to happen. We still need to read a billion Chessex, right? A year. That's not going away. And if you don't solve that, adding new diagnosis and new techniques is not going to fix it. Not going to matter, yeah. But some smart people, and we have a lot of them in the world, need to build on both stack to allow us to have a better health system in the future. So that is a perfect setup for my last question for you. Speaking of the people who need to come along and build on both stacks, for folks listening, maybe younger folks, maybe people making a career switch, whoever they may be, who are interested in pursuing a career in AI and perhaps specific to using AI and other advanced technologies in healthcare, to advance healthcare, what advice might you have for them?

27:21Yeah, I would say the first step is to decide upfront what problems you're willing to spend a huge part of your time solving first before the AI part. So unless you are, So the exception being if you're an AI researcher who your curiosity lies in a discovery of AI techniques and machine learning system. And as a society, we really need those people to keep going because they can probably work. But if the ambition is to apply AI, right? So if the question was about people looking to apply AI, the first thing is actually fall in love with some problem. Because it is hard, right? It is a very hard and long journey.

28:06oftentimes the AI bit, the data science bit might even be the easier part. It's the part where you understand the problem, collect the data set, getting enough resource together to label those data set, hopefully piecing up enough capital to buy some of those fancy media GPU to train those systems. And then at the end of it, it needs to be a product, right? It needs to be something that adds value that people are willing to use and pay for. So it's a very long, hard journey. So unless you really fall in love with that problem space and willing to spend your life best work solving that problem, it's going to be very hard to think about it on just AI front.