Kevin Lyman, "Deep Learning in Medical Imaging"

Kevin Lyman on the Angel Invest Boston podcast.

Gamer turned AI maven, Kevin Lyman is implementing Jeremy Howard’s vision of harnessing the super chips that power gaming machines (GPUs) to interpret medical images. Deep learning algorithms are already making radiologists faster and better at their work. It’s a story of man plus machine, not man versus machine.

Don’t miss this accessible report from the frontier where deep learning is starting to make a difference in medical care.

Click here to read the full episode transcript.

 

Some highlights from the interview:

  • Kevin Lyman Versions 1.0 & 2.0

  • Kevin Lyman, Gamer Turned AI Maven

  • Stints at Hasbro, SpaceX & Microsoft

  • The Inventors Guild

  • Jeremy Howard of Kaggle Fame Decides AI Is Ready to Take on Medical Imaging - Founds Enlitic

  • GPUs: Good for Gaming + Processing Medical Images

  • “…deep learning has managed to turn AI into more of an engineering problem than a science problem.”

  • “…you can see most AI setups today as being like a black box.”

  • Kevin Lyman Dropped Everything to Join Jeremy Howard at Enlitic

  • “…they didn't just give us the $10 million in our Series A. They gave us 100 terabytes of historic patient data-“

  • “…21% faster, they caught 11% more of the true positives and with 9% fewer false positives…”

  • “…we should not be focused on man versus machine… We should be very focused on man plus machine…”

  • Path.AI – Compare & Contrast

  • Kevin Lyman’s Personality Disposes Him to Work in Startups

  • Kevin Lyman 2.0 Interview – Nine Months Later

  • Headway During Nine Months Since 1.0

  • 95% Body Coverage

  • About to Deploy Beta with Capitol Health


Transcript of “Deep Learning in Medical Imaging”

Guest: Kevin Lyman, CEO of Enlitic in “Deep Learning in Medical Imaging”

SAL DAHER: Welcome to Angel Invest Boston, conversations with Boston's most interesting angel investors and founders. I'm Sal Daher, an angel investor curious to learn more about how to build successful new companies.

SAL DAHER: The best way I can think of doing this is by talking to people who've done it, people such as my guest, Kevin Lyman of Enlitic, a startup backed by a major strategic player that is applying deep learning to the interpretation of medical imaging.

Kevin Lyman Versions 1.0 & 2.0

SAL DAHER: In this episode, we will have the good fortune of listening to two versions of Kevin Lyman. Kevin Lyman Version 1.0 is the one I met at TEDMED in Palm Springs nine months ago. He was then COO of the company and he was delivering his TEDMED Talk.

SAL DAHER: In our chat, we learned how this full-time gamer got into artificial intelligence and its applications in medicine. We also chatted about how Enlitic was founded and the merits of collaborating with a strategic player.

SAL DAHER: More recently, we embarked in a conversation with Kevin Lyman 2.0, who is now CEO of Enlitic. That conversation covered the progress the company has made since the 1.0 interview, as well as the story of how Kevin became CEO.

SAL DAHER: Welcome to Angel Invest Boston, conversations with angels and founders. We usually are coming to you from Boston. We are now in Palm Springs, California at the TEDMED conference interviewing people who have presented there and are participating in the conference, and we are very lucky today to have Kevin Lyman of Enlitic.

Kevin Lyman, Gamer Turned AI Maven

SAL DAHER: Kevin has a very interesting background, gamer turned AI maven, working in something that is absolutely fascinating, which is the application of AI to medicine. Kevin, welcome to the Angel Invest Boston Podcast.

KEVIN LYMAN: Thank you for having me.

SAL DAHER: I'm so glad you made the time to be here. You're a very busy man.

KEVIN LYMAN: Very glad that you made the time to meet with me.

SAL DAHER: Kevin, usually we start our podcasts by just asking our guests to tell us how they got to where they are, how they discover that's what they want to do with their lives. I'm not asking about how the company started. How did you get to where you are doing this remarkable thing to the point where you're accepted to the Hive?

KEVIN LYMAN: I have a pretty weird background, probably a very different journey for most people that end up doing things in AI or healthcare.

SAL DAHER: Okay.

KEVIN LYMAN: I actually used to be a professional gamer when I was in high school and played professional Halo 2 and then World of Warcraft, and that taught me a lot about team building and just trying the same thing over and over until you get good at it.

KEVIN LYMAN: As silly as it might sound, it was really a forum for me to learn that if you work hard enough at something, you can be really, really good at it, but a lot of times it's just the determination and the effort that is the step that most people don't put in.

KEVIN LYMAN: At a certain point, I decided to shift that away from gaming and maybe try to apply the same thing to technology and doing bigger things. By the time I'd gone to college, I decided that the best way to do that was to take some time off and work at a bunch of different companies, pick companies that were doing things I found very interesting, and learn everything I could from it.

Stints at Hasbro, SpaceX & Microsoft

KEVIN LYMAN: I went to Hasbro first, the toy company, where I had a super cool job, rapidly prototyping high-tech concept toys and pitching it back to the board to see if it was something worth putting money into, and I just learned a whole lot about working with emerging technologies and rapidly putting together a working thing and selling a vision around it.

KEVIN LYMAN: From there, I went to SpaceX, where I was developing control circuits and sensors for the Falcon rocket and the Dragon space capsule. That's actually where I started doing some machine learning work, doing early simulation models for the rocket landing system.

KEVIN LYMAN: From there went to Microsoft to do some user experience research on Excel, which I thought was a super interesting problem because everybody uses Excel. That's the most widely used software development platform in the world and most people don't even know that they're programming. From a user experience research standpoint it's a super, super interesting problem.

The Inventors Guild

KEVIN LYMAN: I had taken all of those learnings and used it to start a couple companies of my own. The first was a company called The Inventors Guild, which is a team of students around the country that work together to consult for startups and tech companies. Really, that was an opportunity for me to build a big network of very smart young people to work with on all these cool projects.

KEVIN LYMAN: In the process, a couple of startups had sprung out of that, one in the medical device space building wearables for clinical trials, and the other doing machine learning for recruiting, resume analytics.

KEVIN LYMAN: I had been working on both of those things, and one of my co-founders from the Guild had introduced me to Jeremy Howard, who is sort of a celebrity in the machine learning world. He's the former president of Kaggle, the machine learning competition platform.

Jeremy Howard of Kaggle Fame Decides AI Is Ready to Take on Medical Imaging-Founds Enlitic

KEVIN LYMAN: He was putting together this vision of Enlitic, the, "Let's bring the best doctors in the world together. Let's bring the best AI people in the world together. Now that AI is ready, let's apply it to solving this incredibly important problem of diagnostics."

KEVIN LYMAN: I just fell so in love with that vision and it just made so much sense. From the technology standpoint, it was the right time. From the mission standpoint, what could you do that's more important?

SAL DAHER: Let's go back a little bit to what you said about Microsoft Excel and programing. I just want to follow up on that. Are you talking about people writing macros and so forth?

KEVIN LYMAN: Well, macros are a part of it, but even just putting numbers in cells and building formulas on top of that is illustrating the exact same logic that most people are going to put into software development.

SAL DAHER: Sure, yes.

KEVIN LYMAN: Even if you're not writing those macros, you're still exercising that same concept. I've seen some very, very impressive spreadsheets in Excel from people that don't realize that they're actually pretty good developers to have put that together in the first place.

SAL DAHER: Very good. Where are you from originally, Kevin?

KEVIN LYMAN: Jersey.

SAL DAHER: From the New Jersey.

KEVIN LYMAN: Yeah.

SAL DAHER: Where did you go to undergrad?

KEVIN LYMAN: Rensselaer Polytechnic.

SAL DAHER: RPI.

KEVIN LYMAN: I'm so surprised every time people know what RPI is.

SAL DAHER: Of course. Of course, RPI. RPI is a very prominent university. You went from RPI, and then you became a professional gamer.

KEVIN LYMAN: I was actually a pro gamer before that, when I was in high school.

SAL DAHER: Pro gamer before that. This podcast is a little bit the bane of many parents. You might want to say that you didn't make a career in Halo 2. You made a career in AI later on. It led to a career in AI.

KEVIN LYMAN: Yes.

SAL DAHER: Excellent. Let's talk about now the founding of Enlitic and how that came about. What's the founding story there?

GPUs: Good for Gaming + Processing Medical Images

KEVIN LYMAN: Artificial intelligence has been around for a while. It's a very general concept. The aspect of AI that most of us now consider to be AI is a particular field of algorithms known as deep learning. Deep learning, neural networks have been around for a long time, 30 years.

KEVIN LYMAN: It's just very recently, like sometime around 2012, it started to be reasonable to actually use these algorithms. Computational power caught up to what is required to actually use these things. It turns out they're very, very good at imaging analytics, in particular, when you can run them on GPUs or graphics processing units.

KEVIN LYMAN: Somebody had the great revelation that the same cards that enable gaming are also really, really good at matrix multiplication, which is all that you really need to do for this imaging AI analysis.

SAL DAHER: That is too cool.

KEVIN LYMAN: Yeah, it's amazing.

SAL DAHER: You were using the technology, but in a different way.

KEVIN LYMAN: Yes, yeah. It almost feels like it was destiny or something.

SAL DAHER: Where did you pick up all the linear algebra that you needed to be doing AI and all the other ancillary stuff?

“…deep learning has managed to turn AI into more of an engineering problem than a science problem.”

KEVIN LYMAN: Well, I say that I feel very blessed as an engineer who managed to find a way into healthcare. I feel the same way about science and AI in general, that deep learning has managed to turn AI into more of an engineering problem than a science problem.

SAL DAHER: Okay.

“…you can see most AI setups today as being like a black box.”

KEVIN LYMAN: By that, I mean you can see most AI setups today as being like a black box, where on one side you have an X and on the other side you have a Y. This magic box called deep learning can figure out how to get from X to Y, by learning from enough examples.

SAL DAHER: Okay.

KEVIN LYMAN: Now, there are tricks you need to play when crafting that box in order to optimize it to make sure it converges faster and finds the right answer in a reasonable amount of time, but ultimately, there's enough research and understanding of how to do these things that even if you don't know a whole lot of math and statistics, you do a pretty good job.

KEVIN LYMAN: That's really where I started with it. How can I start with something that works, and along the way learn everything I can by working with the world's best AI people and also the world's best doctors to pick up that math along the way?

SAL DAHER: You were sort of self-taught.

KEVIN LYMAN: Yeah. I'm the kind of person that likes to learn by jumping in head first and spend 20 hours a day with the world's best people learning everything I can about it.

SAL DAHER: Right, right. Excellent. How did Enlitic come about? Tell me the founding story there. GPUs becoming more powerful, AI, so the idea is to apply it to medicine.

KEVIN LYMAN: Yeah, so sometime around 2014, machine learning celebrity and former president of Kaggle, the machine learning platform, Jeremy Howard, had this revelation that now is the perfect time to apply machine learning and deep learning to healthcare, in particular, imaging analytics, which was an area that is prime for it because everything is already digitized.

SAL DAHER: Right.

KEVIN LYMAN: At the same time, radiologists are used to doing everything with software, but previously was an impossible problem to tackle because the algorithms just weren't there. The computational power wasn't there. Jeremy started to bring together the best doctors and the best machine learning practitioners to tackle this problem.

Kevin Lyman Dropped Everything to Join Jeremy Howard at Enlitic

KEVIN LYMAN: Around that time, a friend of mine that I had founded a previous company of mine with had introduced me to Jeremy, and I just fell so in love with the vision and the mission that I had to come immediately onboard with it. It was actually within three days of meeting Jeremy that I wrapped up my other companies and moved to San Francisco and said, "This is what I'm doing full-time now."

SAL DAHER: Oh my. Where were you living then?

KEVIN LYMAN: I had actually just moved to Sunnyvale. I moved my startups from New York to Sunnyvale to be closer to our investors.

SAL DAHER: Very good. Tell me the current circumstance of Enlitic. How far along are you?

KEVIN LYMAN: We founded it in August 2014, and so we're just over three years old. We've raised a seed round and a Series A totaling 15 million to date and we're looking to raise again in the next few months. We have developed out a lot of models in a lot of different areas and spent this year really doing a lot of blind tests and validation and trying to be a bit more transparent about a lot of what we're doing. Everything up to this point has really been building infrastructure and understanding. Both of those things are critically-

KEVIN LYMAN: ... infrastructure and understanding. Both of those things are critically, critically important in AI, especially in health care, because we're not out there to build out one diagnostic test. We don't want to just build chest x-rays. We want to do all of diagnostics and we want to do it in a way that makes sense, where we can deliver value along that path. Doing so means you need to understand what goes into building all kinds of medical AI, what is an appropriate process for doing that, how can you share representations among different problems that you're tackling.

KEVIN LYMAN: So, we've put together a platform that enables us to serve medical data that we've collected from our partners to a team of doctors that we have assembled around the world who used that platform to annotate the data for us, to train the machine learning models. They annotate it using custom guidelines that I developed alongside our chief medical officer by consulting with key opinion leaders in different diagnostic specialty areas around the world. In the process, we haven't just built that platform. We've built an understanding of how do you figure out the actual thought process of a doctor and translate that to something understandable for an AI, which turns out is a fairly complicated problem-

SAL DAHER: Oh, yeah.

KEVIN LYMAN: ... that a lot of people haven't thought about. They tried to brute force it.

SAL DAHER: Kevin, I understand that your capitalization table is unusual in the fact that you have, as your largest investor, a strategic investor. Could you tell us a little bit about that?

“…they didn't just give us the $10 million in our Series A. They gave us 100 terabytes of historic patient data-“

KEVIN LYMAN: Our largest investor and biggest partner is Capitol Health in Australia. They're the second largest radiology provider. They have really taught us the value of working with strategic investors in that they didn't just give us the $10 million in our Series A. They gave us 100 terabytes of historic patient data-

SAL DAHER: Wow.

KEVIN LYMAN: ... to really form the foundation of everything that we build and access to all of their radiologists who are crucial, not only in helping form our annotation guidelines and our AI strategy, but also in testing those solutions and actually using them, helping us to understand how will this help a doctor in the field.

SAL DAHER: Oh.

KEVIN LYMAN: They have really taught us the tremendous value with working with strategic investors. Really, I can't thank them enough because they didn't just give us money. They gave us 100 terabytes of historic medical data to really start us off on everything and access to all of their radiologists who, on an ongoing basis, help us do everything we do.

“…21% faster, they caught 11% more of the true positives and with 9% fewer false positives…”

KEVIN LYMAN: Of course, we test our solutions with them, too. We actually ran a study with them last year, which was the first clinical deployment of deep learning where we used one of our models for a fracture detection to automatically circle fractures in x-rays. We integrated that with the software they were already using. As far as the doctors are concerned, they didn't necessarily need to know if an AI was circling everything for them, or if it was just another person that looked at it before them. We ran a study to see how much better were radiologists augmented by the system versus those who are not. We found that the people using it were 21% faster, they caught 11% more of the true positives and with 9% fewer false positives, which is pretty great. Ultimately, we decided this is a good proof point that this can work. This can help radiologists do their job.

KEVIN LYMAN: That's when we started to focus on more clinically-impactful problems, one of which was x-ray interpretation makes up 45% of all radiology images, or other studies, I should say. The other being chest CTs, where we focused specifically on lung cancer detection and diagnosis. Capitol and several of our other partners have been massively helpful in testing both of those solutions.

KEVIN LYMAN: We actually have three different ways that we can deploy those models. The first is for quality assessment. So after they write a report, we use an image model to look at the image and a text model to independently look at the report they wrote about that image and we see if it matches. Did you miss something? We do that as an entry point because, from a regulatory standpoint, it's quite a bit easier. It's more of an administrative tool than a diagnostic aid. From a radiologist's standpoint, it's clearly not trying to steal their job. It's trying to help them. It's trying to make sure that you didn't miss anything in that image.

SAL DAHER: So, you're improving the performance of the radiologist. That's the plan for now, right?

KEVIN LYMAN: Mm-hmm (affirmative).

SAL DAHER: So, like the base camp. Summit is eventually to get approval to be just complete AI, producing the reports and everything.

“…we should not be focused on man versus machine… We should be very focused on man plus machine…”

KEVIN LYMAN: I think we're a long, long ways away from that, and I would be very skeptical of anybody that tells you otherwise. I think right now we should not be focused on man versus machine, at least in health care. We should be very focused on man plus machine in health care. I think it's going to be that way for at least a few decades.

Path.AI – Compare & Contrast

SAL DAHER: Very good. Are you familiar with a company called PathAI?

KEVIN LYMAN: I am familiar with PathAI.

SAL DAHER: Compare and contrast what you're doing versus what they're doing.

KEVIN LYMAN: Path.AI is focused more on pathology imaging, as their name suggests, and we're focused on general medical diagnostics, but really where we've started out is in radiology. So while we have done a bit of pathology, it hasn't really been our core focus up to this point.

KEVIN LYMAN: Most of what we've been focused on is radiology, but in a way where we can generalize that to all different kinds of medical diagnostics so how do we build one platform that can do radiology, cardiology, ophthalmology, oncology, etc.? Because it turns out that building all of those things together actually makes a pretty big difference in terms of the insights that those things can share.

Kevin Lyman’s Personality Disposes Him to Work in Startups

SAL DAHER: Excellent. Now, Kevin, you are a young man with skills that are extremely valuable. What convinced you that the best way to use those skills is in a highly uncertain startup, instead of just going into Google or in Amazon, any of these places that would be very glad to pay you for your excellent skills?

KEVIN LYMAN: I think different people thrive in different kinds of environments. I'm the kind of person that needs a bit more freedom than you can really get in those of sorts of environments. I think that large companies like Google and Microsoft and Amazon are a great place to learn a lot. That's why I really wanted to take advantage of those things to learn everything they had to teach me, or at least a lot of what they had to teach me while I was still in college.

KEVIN LYMAN: But, ultimately, that's how I realized that I'm the kind of person that needs to be in a startup. I think your propensity to push change is just so much greater. In a company like a Google or a Microsoft, I would not get the opportunity to interview doctors all day and night, learning everything they know about their diagnostic area. I wouldn't have the opportunity to speak at TEDMED about this incredible mission.

KEVIN LYMAN: Probably the most important thing to me, I wouldn't have the ability to actually make that change that really drives me. That's really what it is about for me. I want to push really great technology and I want to make a big difference for everybody in the world. I think this is the best way for me to do that.

SAL DAHER: That is tremendous. Kevin Lyman, thanks a lot. I'm very grateful that you took the time to sit down and be with us here.

KEVIN LYMAN: Thank you very much for having me. I thought it was a really great conversation.

SAL DAHER: This is Angel Invest Boston coming to you from Palm Springs, California at the TEDMED conference.

Kevin Lyman 2.0 Interview – Nine Months Later

SAL DAHER: Hey, Kevin. It's good to be talking to you again. Nine months ago, when we last spoke, you were chief operating officer, COO, and lead scientist of Enlitic. You are now CEO. Please tell us a bit about that journey.

KEVIN LYMAN: I think it's a really interesting story and I'm very grateful in a company like Enlitic to have worn all the different hats, even prior to being COO and lead scientist.

KEVIN LYMAN: I originally had joined the company as an engineer, then benefited quite a bit in being one of the few people to get to travel onsite with partners and really spend a lot of time upfront in the field with radiologists, really understanding exactly the problems that they encounter on a daily basis that highlight the need for this type of technology. From there, moved directly into the modeling side of the business, working very closely with our other data scientists and, again, with our clinicians to capture various insight and translate that into real deep learning work.

KEVIN LYMAN: Progressively, throughout my lifetime with the business have gotten to experience more and more different functions which, in our case, what we're building is so incredibly nuanced and that it requires understanding from so many different angles that really, I get to benefit from insights into the product and technology that made me very grateful for having that kind of viewpoint on it.

KEVIN LYMAN: Over time, it just sort of made sense to shareholders, to the board, to our partners and the other employees in the company that having me in this type of role, being in the CEO position to explain the product from so many different angles, to really be able to empathize with the customer directly and explaining how these systems work and why they're so beneficial that it really just made a lot of sense, then I think it's a really interesting scenario to have the somebody move through all the different positions like that, but one that I think in our circumstance is exactly the right move.

Headway During Nine Months Since 1.0

SAL DAHER: Kevin, since we spoke nine months ago in Palm Springs, what's the headway that Enlitic has made? Please outline for me a little bit the progress that the companies made in that time.

KEVIN LYMAN: Absolutely. It's been an extremely exciting several months since then. We've taken a lot of time to really focus on our commercial model and work very closely with a lot of our partners to set up some very exciting opportunities for us going forward. Largely, our activity right now is around moving forward in two paths. I think last time you heard me talk quite a bit about investing in infrastructure and how critical it is toward building out scalable solutions in this space and we've really taken that to heart. Of course, we've continued investing there, but we'll, of course, continue to do so much more.

KEVIN LYMAN: Largely, a big part of our vision now is building out as much coverage of the human body as possible. A big part of our activity is moving toward attaining 95% body coverage over the next three to four years, aiming to really leverage that infrastructure as much as we can. So that's an area where we're moving forward quite rapidly.

KEVIN LYMAN: On the other hand, we're also moving forward on the commercialization of some of our more advanced systems that we've built out today, moving forward over the next several months on the regulatory approval and subsequent commercialization of our work in lung cancer screening and chest CT, as well as various applications of our work in chest x-ray interpretation. I think both angles there are very exciting for us.

KEVIN LYMAN: Right now, we're going toward building the team out quite a bit. Right now, we're a team of about 16 engineers and data scientists. We're going to be building that out quite extensively over the next several months, looking to add an extra data scientist and an extra engineer.

KEVIN LYMAN: Looking to add an extra data scientist and an extra engineer every month for the next 12 months or so. And we already have many of those candidates in the pipeline accepted and coming on board in the next couple of weeks, and so we're just very, very excited to rapidly accelerate toward both commercialization and a more concerted build out effort.

95% Body Coverage

SAL DAHER: Kevin, please expand a little bit on what you mean by 95% body coverage.

KEVIN LYMAN: Yeah, a big part of our goal right now is to build out 95% body coverage. And by that I mean across the four major modalities of X ray, CT, ultrasound and MRI, building models that can understand each part of the human body, each diagnosis or pattern that a radiologist might detect in those various anatomic regions and even building out a comprehensive roadmap of how one can build out those systems has required going back to square one and understanding the fundamentals of how a radiologist looks at these problems than that.

KEVIN LYMAN: In that, let's look at a chest X-ray study, for example. To a radiologist, it's just a chest X ray study. But to an engineer, to myself for example, looking at it with no radiological backgrounds, I see five different types of images that may or may not be served in any combination to a radiologist. But the radiologist just ignores that concept. They just know it as a chest X-ray study. In reality, it can be an AP a PA, which are frontal and reverse views. It can be a lateral from either direction, or it can be an oblique view looking at either half of the rib cage.

KEVIN LYMAN: And understanding that is critical because when training a deep learning model, you need to understand that what you're really looking at is probably five or six different variants of a deep learning model. Each one trained to look at each of those different images. But to really understand that, we've had to spend a lot of time with radiologists understanding the fundamental building blocks of their job and how that differs from a machine interpretation to a human interpretation. We've extended that out toward again, every type of study that a radiologist does. Looking at every type of X-ray they read, every type of MRI, every type of CT that they read.

KEVIN LYMAN: Rather than looking at that as different types of studies, we've looked at it as different types of views, different types of fundamental imaging protocols. And after breaking all of those different types of studies up into again, those elemental building blocks, we've then reconstructed it back at a macro scale to each type of study that we aim to build out. So, that now we're approaching it, not just from what's most intuitive to radiologist, but what's most intuitive to a team of engineers building this out for a group of radiologists.

KEVIN LYMAN: And we've deliberately structured that in a way that's very modular. Sure we want to build out full interpretation of a spine CT, but we can choose to do that before or after we build out brain MRI. And we can choose which order to build that out and based on what delivers the most immediate value to the partners that we're working with in order to build those systems out.

KEVIN LYMAN: Translating that same kind of thinking to a more specific diagnosis, again, let's look at a chest X-ray, the models that we're building out right now for chest X-rays are trained to detect every kind of abnormal pattern that can occur in that type of scan. But again, that's really required thinking about it, more or less, what steps does a radiologist perform, whether or not they know it to interpret this type of scan. We don't want to just go directly from image to diagnosis, because there's a lot of steps in that process. And we're very likely to miss a lot of that nuance, and either bias our model or overfit it to a data set we don't understand. If we don't spend ample time really determining what problem we're solving in the first place.

KEVIN LYMAN: And in this case, I really liked the point that the example of tuberculosis detection in a chest X-ray, because we see a lot of researchers in the space will try to train a model to get diagnostic accuracy for TB. The problem is that you can't really diagnose tuberculosis from a chest X-ray. The best you can really do is detect signs of it, that will indicate that you should run another test to really diagnose tuberculosis. But because we've looked at it from the perspective of step by step, what mental processes are radiologists following, we know that what they're really doing in the scan is looking for patterns. They're looking for, again, those signs that indicate that there might be another test that they want to run. Now, in the case of looking for TB, that'll mean that the patient might have pleural effusions, cavitation, calcified long nodules, consolidation.

KEVIN LYMAN: These will all give the radiologist some signal that the patient might have TB, but in reality, they still need to look at the patient's medical history to understand where they've been spending time. Have they been coughing? Have they had any other kinds of infections recently. And so coming to that ultimate conclusion that they may or may not have TB, requires more information than just looking at that scan. And so that's just the type of insight that has really required working very closely with radiologists to understand, because otherwise, we might be training a model to do a task that ultimately would be impossible to ask it to perform.

SAL DAHER: Kevin in the 1.0 conversation that we had, you talked a great deal about the importance of a strategic relationship such as the one with Capitol Health in helping to train your artificial intelligence and deep learning engine. So please tell us how things have progressed with Capitol Health since then.

About to Deploy Beta with Capitol Health

KEVIN LYMAN: The main thing that I point to that's new on the Capitol side is that we're moving very quickly with them toward a wider scale deployment of our system. Last we spoke I had mentioned a bit about studies that we've deployed with them in the past and some of our activity there. But as we start moving again toward full body coverage, what we're really doing is working with some of our very close partners like Capitol to deploy a wider system and a wider platform where month to month, we'll be updating it with new types of models covering more parts of the body. And that's what we're currently in the process of transitioning toward, with Capitol. Taking some of the previous deployments we've done there and some of the infrastructure we have in place and tuning that more toward this ongoing beta deployment where you'll see us testing a lot of different types of systems with Capitol and on a very rapid basis.

KEVIN LYMAN: So, we're quite excited about that as well. Aiming to probably have that deployment in place in probably about the next 60 days.

SAL DAHER: Kevin, I'm really glad to hear that things continue to go well with Capitol Health. This is really promising. It sounds like you are accomplishing the things you talked about nine months ago in what you're doing now. So I'm really grateful that you may time in your very busy schedule to be with us and I hope that in the future, you come back and let us know further progress from the Enlitic.

SAL DAHER: Well, listeners I hope you've enjoyed this conversation as much as I have. This is Angel Invest Boston. Conversations with Boston's most interesting founders and angel investors. I'm Sal Daher.

SAL DAHER: I'm glad you were able to join us. Our engineer is Raul Rosa. Our theme was composed by John McKusick. Our graphic design is by Katharine Woodman-Maynard. Our host is coached by Grace Daher.