Ep 13: PathAI CEO Dr. Andy Beck on the Future of AI in Medical Diagnosis

Jacob sits down with PathAI Ceo Dr. Andy Beck to discuss the state of AI adoption in diagnosis, why Path acquired their own lab, pathologists' jobs in the future and nailing GTM to reach a ~$1B valuation.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:02Jacob Effron?

  This is Unsupervised Learning, Red Points AI Podcast, and I'm Jacob Effron, one of your co-hosts. Today, we had a great conversation with Dr. Andy Beck, the CEO and co-founder of Path AI. Path's doing some fascinating work to take pathology data to improve diagnoses and treatments. The company was valued close to $1 billion recently, and Andy had a really interesting perspective on the state of AI adoption in pathology.

• 0:21Speaker 01

  We think there will be a quick adoption because it will make pathologists more accurate, more reproducible in the assessment of biopsies. Similarly, on ecology, we think it's going to be important for adoption to focus on things that pathologists are struggling with and that are critically important for treatment decisions.

• 0:33Jacob Effron?

  Nailing go-to-market.

• 0:34Speaker 01

  Our first go-to-market was drug development, working with biopharma for a number of different reasons. But one is, it was just really good product-market fit between what they need from their completed clinical trials, particularly initially, and what we can provide.

• 0:47Jacob Effron?

  And why Path acquired their own lab.

• 0:49Speaker 01

  We acquired a large anatomical pathology laboratory in 2021. We also set up a biopharma laboratory there, so now we can really service every set of customer.

• 0:58Jacob Effron?

  I think you'll really enjoy our discussion. Andy, thanks so much for coming on.

• 1:01Speaker 01

  Great. Thanks so much for having me.

• 1:03Jacob Effron?

  Well, to start, I'd be curious, you know, obviously, you have no shortage of educational credentials. And as you were kind of navigating the early parts of your career, you know, wondering how you decided to go into pathology in the first place? And obviously, you were early to AI in pathology. What kind of got you interested in that?

• 1:17Speaker 01

  Yeah, no, it's a good question. So I guess, so for some personal history, my mother's a dermatologist and also trained in dermatopathology. So she took care of patients with skin conditions and also kind of like a third of her job was doing pathology. So seeing those things biopsied and looked into the microscope. So I thought like that seemed super interesting. And I remember even visiting, visiting her at work. And it's kind of magical, like the microscope and looking at these images and like one person knows exactly what they're looking at. It's like the key information for each patient they're looking at, like the other person's completely. Kind of like a random set of images that you have no idea what you're looking at. So I thought just this idea of interpreting images being key to guiding medical care was really interesting. And then kind of go ahead to medical school.

• 2:00Speaker 01

  I think the early years of medical school pathology, at least when I took it, played a really important role. Textbooks like Robyn's Pathologic Basis of Disease, I thought like brought all of this stuff together into one field. Genetics, immunology, cancer biology, cellular basis of disease. Like it kind of all comes together and seemed extremely rich and interesting and something that could continue to be interesting going forward. And then, yeah, it really just fulfilled that in the clinical years. The idea of like looking at the primary document for each patient about what's really driving their disease, where it's both kind of scientifically gratifying, but also is one of the most important pieces of information kind of that a physician can deliver for guiding that patient's care. So it's kind of those combination of factors. I actually went to it pretty early.

• 2:43Speaker 01

  It's one of the first things I really latched on to during medical school and have kind of been doing it ever since.

• 2:48Jacob Effron?

  Yeah, no, that's awesome. And then obviously I feel like you were, you know, in the early days of actually applying kind of algorithms to, you know, pathology. You know, maybe it's a good introduction to kind of just giving a brief context of like the history of providing more like quantitative methods on top of some of the qualitative work within pathology. And I would love to hear kind of your journey through that to ultimately starting Path AI.

• 3:09Speaker 01

  For sure. So one thing that was always super interesting about pathology, and I think every patient can relate to this, there's two big wings of it. There's clinical pathology and anatomic pathology. And each of us have certainly had results come back from each one. So you're thinking pathology broadly is kind of like the laboratory basis of medicine. So something like 60 to 70% of medical decisions, and this is a big like $4 trillion market, are guided by some sort of laboratory result. Clinical pathology is largely blood-based work, and that over the years has been highly centralized, highly made objective, reproducible, standardized across the US and across the world, so that when you get, you know, a glucose level or an LDL cholesterol at one

• 3:54Speaker 01

  lab, it's a quantitative number. And, you know, if you give that exact same sample to another lab, they have good ways of either making sure that's the same or identifying, you know, when is it out of range? Then you take the other half of pathology, which is anatomic pathology and largely what I've focused on both in my training, but also at Path AI. That's this like very almost artistic expert field that truly hasn't changed a lot in the practice of it in many decades. And that's where if you go to the doctor, and let's say they do a blood test and you're a man and you're 65 years old or something, and your PSA is elevated, your prostate-specific antigen, that's a quantitative test, but they're going to want to know, well, do you have prostate cancer? The only way to know that, to confirm the diagnosis, and this remains true today, it'll

• 4:38Speaker 01

  be true with blood-based markers and everything else, is you take a tissue biopsy, and you take this physical piece of tissue out of the body, you embed it in a wax block, you cut thin sections out of a microtome, you have these kind of five micron sections that are put under a microscope, and it really is a human today using our powers of perception to interpret what is going on in this image, to make critical diagnoses like, do you have cancer, yes or no? And if you do, what's the grade and what's the severity of it? So that's very much the area of pathology we've worked on, because, one, it's incredibly important for diagnoses of diseases like cancer, inflammatory bowel disease, and NASH; and, two, it's really difficult to do objectively, reproducibly, scalably, as well as as precision

• 5:20Speaker 01

  medicine continues to advance, to continue to advance that field, to keep pace with the rate of progress in therapeutic development. So, one, there's a big problem, you know, a big opportunity here, I would say, because it is so important. And we've had this incredible transformation over the past decade in, you know, computer vision. So I would say, to back to your question, when did I start? I started very early. It was actually one of the first projects I worked on at medical school. We were doing a neural net back in essentially 2004 or so, so like 20 years ago. So I saw how bad it was back then. And the vision's been there for 50 or 60 or 70 years, like people have been trying to do this stuff. I would say 20 years ago I was trying, but it was so far removed from what a human could do.

• 6:04Speaker 01

  It was so artificial and costly. And then kind of I've been following in the field for 20 years, and yeah, the last, you know, five to 10 years has just been incredible.

• 6:12Jacob Effron?

  And what he said, I mean, obviously, you've done a lot of academic research around, you know, implementing these algorithms. Like, when did you decide, hey, like, this makes sense to the company, and I'm going to go start, you know, Path AI?

• 6:21Speaker 01

  Yeah, it was a couple of things. So, one, I had been in the field for a while. So, I'd say we started the company in 2016. So, I'd probably already been thinking about these problems for like a decade and had seen kind of where they were, what was state of the art before deep learning. And then, you know, the big advance in deep convolutional neural nets in computer vision were around 2012. And this is a very different application area, but just sort of showing that, wow, this is incredibly better than anything before. And then, it wasn't a fake. It was like, year over year, these systems were getting incrementally better, to the point where, for many applications, around the 2016 timeframe, computer vision systems were outperforming humans for similar tasks. And this just like hadn't been true for, you know, if you look back a decade, the joke was always a three-year-old

• 7:04Speaker 01

  is better than a system trained by like, three dozen PhD scientists with hundreds of millions of dollars. And that was very disheartening, I think, to some people in the field. That all changed around the 2015 timeframe. So, that was one big thing. And the other was just a couple of other big sort of secular forces, like cloud computing was a huge advance. And then, digital pathology overall, it was clear that that was going to mature and ultimately gain widespread adoption. So, not just creating images on a microscope that eventually, you know, are just gone once you're done looking at them, but actually taking a big picture, creating a permanent inexhaustible digital resource of every image, so that a human can analyze it, but also, in the future, algorithms can analyze it. So, it was really those three things, increasing availability of digital

• 7:47Speaker 01

  data through digital pathology, the incredible advances in deep convolutional neural nets, and cloud computing. And then, it was just a good time for me to really, kind of, I was at a point in my career where I wanted to try something new, and, you know, saw this big problem that could be solved and a big opportunity. So, that was kind of the personal and the sort of things in the field that came together for starting the company when we did.

• 8:06Jacob Effron?

  Totally. And then when you got started, I mean, I imagine, you know, to your earlier point, I mean, there's so many different conditions that an atomic pathologist work in, you know, whether it's cancer, inflammatory bowel, NASH, like, how did you kind of decide where to start?

• 8:18Speaker 01

  Yeah, no, it's a good question. I would say one thing we started and still an important theme in the company is what's not going to change? What's robust to support all of these end use cases? And kind of when thinking about a company and applications and building a platform, you kind of want it wide enough in scope, but kind of strategically to be narrow enough that you invest heavily and triple down on something and then, you know, it will hit a wide enough scope. So you don't want to make the market too small. You obviously don't want to try to boil the ocean. So the good thing here is there's something that's very static and robust across all these applications, and that's these whole slide images of tissue that are a certain size and have certain characteristics that are used and used by pathologists for all these different conditions. So one is

• 9:03Speaker 01

  really thinking about a technology platform and user facing interfaces that will cover the full spectrum as like a first area to invest in. So not just focusing on specific applications, like we're going to go after early stage year positive breast cancer and throw all of our attention there. And that may not at all be transferable to, you know, late stage or lung cancer. We wanted to focus on what's going to cover everything where tissue histopathology is important. And that included the technology platform, the interface for research, for clinical trials, for clinical practice, as well as the ability to generate very large, well annotated training data sets. And of course, the machine learning capabilities to predict what are the cells in the image, what are the tissue in the images, can I predict molecular phenotypes in the images, can I predict patient outcomes, treatment response

• 9:47Speaker 01

  from the images? All of that, those methodologies are very general across those use cases. And then we really did focus, okay, now that we've got this platform and machinery, you know, how do we prioritize which algorithm products to go after? And I can talk more about that. But that's really how we thought about technology and platform first, and algorithm products are prioritized after that.

• 10:06Jacob Effron?

  You know, it makes a ton of sense. I mean, I mentioned there's a lot of common workflows that you'd want to build across the board for pathologists. And then also, you know, just, just kind of integrating with the set of tools they're, they're using in the first place, you know, having the ability To surface focus on the trial side. You know, that common set of infrastructure makes a ton of sense. You know, I imagine when you were deciding, you know, initial use cases go after. A big part too was, you know, just where training data is available, right? I'm actually curious, I'd imagine, you know, when you were starting the company in 2016. My impression at least is there wasn't a ton of this that was digitized, right? And there wasn't like these, these massive data sets that already existed. But how did you go about even getting these training data sets in the first place?

• 10:40Speaker 01

  Sure, I would, I'd say it's an interesting, it could, that could have been the way to do it. But I would say sometimes I think that can mislead you. Cause then, you know, it could be data sets for a problem that's just not valuable. And I do think that's happened. And I would say we've tried, we've tried to prioritize with a different framework and then really focus on getting the targeted data sets that are needed to solve that problem. Versus sort of saying this data happens to be available and using that to drive the application. And I think the biggest things we've thought about.

• 11:09Speaker 01

  So one, something that's truly actionable, where the pathology diagnosis and the precision diagnosis really makes a difference. Cause actually a lot of things just don't matter that much. For example, imagine you have a small tumor on your skin and it gets cut out and there's no other treatment option and you're definitely going to cut this thing out. Because it's just that what the pathologist says is useful in some way, but it's not critical because the procedure is already being done. You're, you're not going to just leave it and you're not going to give chemotherapy or something. So it's not actual, so it's focusing on where the pathology really is the critical information for guiding important medical decisions for patients. So that's one. Number two, where pathologists are struggling and where there's an opportunity to, to really improve in a significant way where you could actually then cause a change in the way that workflow is happening today.

• 11:55Speaker 01

  And then tied to two is some evidence that machine learning will effectively solve this problem. And then third, kind of prioritizing based on the markets that will address the most number of patients and really have the biggest impact in both drug development or in diagnostics. So that's kind of the framework. And once you do that, actually targetedly getting data. Is the easier problem to some degree because there are one big advantage of our platform is it's based on whole site images coming from glass slides which are being created at large scale every single day for you know patients undergoing routine procedures across the US and across the world. So it's really asking the questions, having the vision of what the product is in mind and then target going after that data. And of course there's multiple types of data that are useful. There's the images, but then there's sort

• 12:39Speaker 01

  of the annotation from experts. And we focused heavily on both and really building large collections of data, but also a very large set of expertly trained pathologists connected through our platform to help annotate those images. And we use both of those for training our models.

• 12:55Jacob Effron?

  Yeah. And to start, did you have to like, you know, basically find some provider partnerships or pathologists that were willing to kind of share their images with you? And then you kind of built a network of expert annotators?

• 13:04Speaker 01

  Exactly. Yeah. So we have, we, and starting from the beginning, we started with partnerships with academic medical center laboratories, independent laboratories, hospitals, as well as about two years ago, we acquired our own laboratory that also has a large source of data. So all of those are valuable sources of data to help train these models.

• 13:25Jacob Effron?

  Yeah. I want to get into that because I think it's a super interesting, you know, thing that you guys did, you know, in acquiring your own laboratory. You know, I imagine like when you started the business, there were so many different markets, go to markets you could have gone down, right? You know, there was obviously the opportunity to, you know, sell directly to, you know, providers and systems. There was the opportunity to sell to pharma, you know, the opportunity to sell to insurers that want a better care. And then, you know, there was this also, I imagine, like, do you partner with labs? Do you acquire, you know, and it seems like you guys have kind of dual, you know, I think the answer is probably that your go to market is a combination of a lot of these things. But I'm wondering, you know, where you initially started and how that's kind of evolved over the course of the business?

• 14:02Speaker 01

  For sure. Yeah. And I think, as you mentioned, for all these, timing is so important and kind of knowing, controlling what you can control, controlling market development that you can control, but also working with The adoption curves that are external to the company and really timing go to market. And it's not just adoption, it also has to do with reimbursement timing as well as regulatory timing for, you know, putting all these pieces together to how do you really continue growing sustainably over time? And I would say that was probably the topic we thought of most and the biggest strategic question, certainly in the early days of the company. And we did consider all of those options at the beginning. We considered, do we just want to be kind of the intel inside for large diagnostics companies, which is another one you didn't mention, but that was, that was one of the

• 14:47Speaker 01

  initial considerations. Do we want to provide this sort of technology-enabled service to biopharma? Do we want to set up our own lab, just ourselves from scratch, kind of a startup lab? There are many startup companies who do start labs. Usually it's for some kind of specialty testing, but we could have gone that route and then marketed directly to physicians, or we could have been a technology platform to sell into labs. And the kind of, we started first, and I would def say our first go-to-market was very much, and still remains a super important, central part of our business, is drug development and working with biopharma. It's for a number of different reasons, but, but one is, it was just really good product-market fit between what they need from their completed clinical trials, particularly initially, and

• 15:31Speaker 01

  what we can provide, and the ability for them to have, you know, every, every large top 20 major pharmas running hundreds of clinical trials and spending billions, you know, over 10 billion for certain large companies on research and development. So it's incredibly important that they understand why these patients and trials are responding or failing to respond to therapies, to get the most value out of these trials and to help them provide the basis of data for designing either later-stage trials for programs that are continuing, or new combinations or new sets of trials. So that was one where there was clearly a real need and that our technology could really deliver, because it is fundamentally a machine learning problem, is understanding, you know, we have this set of data where we have treatment on every patient, outcome on every patient.

• 16:15Speaker 01

  And pathology images on every patient, how do you use the images to predict the treatment response and how can you use that to identify novel biomarkers or novel patient subsets? So that was a really great place for us to start. It remains a very important part of our business. And that was one that, you know, we could really start scaling relatively quickly with, you know, solving problems for our first partner and then extending that to additional partners. And it's also a very concentrated market, so you don't need a, you know, a huge distributed sales force to target a large portion of the market.

• 16:43Jacob Effron?

  I imagine you're kind of in the, in the early stages of clinical development there. And so, you know, it's possible as you identify these biomarkers, well then you guys become part of just a companion diagnostic or standard of care ultimately for those, for those therapeutics. And then, you know, I also imagine it, it, obviously like to your point, there's much more centralization on the pharma side than it sounds like there is on like the anatomic pathologist side. And so, you know, if, if the pharma companies are getting this data anyway, it's a really interesting way to just be able to, you know, find a few of those versus, you know, did you guys try, like, you know, what's the kind of effort been like distributing to the, you know, maybe more fragmented side of providers out there?

• 17:17Speaker 01

  Yeah, so I'd say that, that was the first big focus of the business and it remains a very important part. We've extended from sort of the exploratory translational research to being able to deploy our systems prospectively in clinical trials as well as to support companion diagnostic development. So that was kind of even within pharma, the, the time course of where we focused. And now, you know, for companion diagnostics to truly be successful and transform medicine, they have to be widely available in this highly fragmented distributed market. And the technology we're building off of is widely available and highly distributed. The histology equipment to create the slides, digital pathology scanners are being more widely distributed. Luckily, the cloud is, is also widely distributed.

• 18:00Speaker 01

  So it really is sort of supplementing all of that and adding in our image management system, our viewer and our algorithms. And to really, I mean, that's the core of the strategy is, is both in that they're very synergistic. You're distributing novel algorithms into practice and then the models over time will get better and you'll acquire more data from many more patients and then that feeds into even better services to advanced drug development.

• 18:24Speaker 01

  And I would say our foray into that market was one, and we've done this with both pharma, with drug development and in diagnostics, is to have both a centralized and a distributed model to take advantage of the advantages of both. And we acquired a large anatomic pathology laboratory in 2021. We also set up a biopharma laboratory there. So now we can really service every set of customer, ones who are already digital through just giving out our digital platform, ones who have the capability to make glass slides, who they can send their glass slides to us to digitize. And even if they don't have any histology capabilities, we can receive tissue to process histologically, digitize, and then apply algorithms. So really both on the drug development as well as diagnostic side, having both this central

• 19:06Speaker 01

  lab model, but also distributed technology platform model allows us to have access to the full market. And it was really just this year that we've started to deploy our technology platform called AI site into labs and hospitals around the, folks starting in the US, but now we're actually doing a lot of work in Europe as well, ultimately expanding that to rest of world. So it was a lot about timing, but ultimately, I think both is the winning model. There's a lot you can do yourself and really optimize things internally and then distribute through the distributed technology platform widely.

• 19:39Jacob Effron?

  It sounds like AI site, you know, works well, you know, when you can get pathologists to adopt the set of things you need to digitize, you know, the inputs. So just to go into a bit more detail there, like what do you actually, from like a behavior change or like equipment change perspective, like what do you kind of need a pathologist lab to have in order to, you know, be able to just kind of provide the software and algorithms on top?

• 19:59Speaker 01

  Yeah, I mean, the beauty is they don't need, because of the unique setup we have, we can service any customer who's, any pathologist who's interested in using the AI site platform. And on one extreme, they're in a lab that's already digital and already has basically the histology equipment, the staining equipment. They also have a whole site imaging system, and then that image can be uploaded into AI site. But the other extreme is just say a pathologist who's operating, you know, in, who's operating by themselves, say, in their office, and they don't even have histology equipment, but they're connected to a physician's practice. That physician can send tissue to our own lab. We can do the histology, create the glass, do it digital, and then serve it through.

• 20:44Speaker 01

  Through the web, back to this pathologist, wherever they're sitting, and they can access AI site in their office and sign out there. So we can really offer the service to the full set of pathologists, from fully digitally enabled to not digitally enabled at all, beyond having a monitor and an internet connection.

• 21:02Jacob Effron?

  And so is part of the impetus behind that, like basically the idea that, look, it might take a long time before the vast majority of pathology offices adopt this digital equipment?

• 21:11Speaker 01

  Absolutely, yes. So if it were like, you know, as mature as, say, like, you know, cell phones or something that are widely distributed, and you just like update it up and push it out to everyone, yeah, I think it would be the model, it would definitely be a different market and a different potential timing around the go-to market. There's still advantages to operating the end-to-end workflow in a single site in terms of the research and development side and fully optimizing the full workflow beyond the algorithm. You can also do that through partnership, but there are advantages to controlling the whole process internally prior to the distribution.

• 21:49Jacob Effron?

  And so will you take, I mean, obviously, part of the model is that folks can send in glass slides, but will you also take kind of like initial reads yourself? Like, do you guys end up being like a referral partner for folks that refer out to pathology? And is that like, I mean, do you kind of compete with some of the pathologists you also support?

• 22:02Speaker 01

  That's a question that comes up. I mean, the short answer is no. Like, we have, we really want to support, you know, all the physicians that are sending the specimens, regardless of where they're sending it, whether they're sending it to us or sending it to other labs. I mean, in theory, you could say some of the pathology labs, there could be some physician practices that we're both trying to get business from. But truly, we want to either enable them or do the end-to-end work ourselves. And we've found, you know, the market is very large, very distributed, so we really don't see that interaction.

• 22:37Speaker 01

  Happen very often and our goal is really to just really support, you know, advancing the field and use of the technology so that, you know, AI site and our algorithms are most widely used regardless of whether it's through our own lab or through a partner lab.

• 22:49Jacob Effron?

  That makes a ton of sense and I guess, you know, I'm sure something you guys thought about a lot from the beginning is just how these algorithms, algorithms interface with, you know, the pathologists themselves and kind of the human in the loop versus, you know, pure diagnosis model and I think probably a good, you know, introduction to just talk a little bit about how regulation works for these algorithms today and, you know, how you guys thought about like the right way to structure, you know, what you guys do to fit into that.

• 23:13Speaker 01

  Yeah, no, it's a great question just because you said how it's a spectrum and there's no kind of at all one size fits all and the algorithm itself is even just one part of the overall device that you're considering. So we very much think about what is the context of use, how is this being used? To your question, what's the role of the pathologist, what are some of the, the risk mitigators if the system is making mistakes? Is there a special control, which in many cases would be a pathologist? And we follow the, the regulatory path that fits for each one and we really do the full spectrum. So a lot of our exploratory research is, is sort of fully automated systems in the sense that the input is whole site images, the outputs are, you know, spreadsheets full of data that then are used by the scientists.

• 23:58Speaker 01

  And there's, there's QC and quality control, but in general, not every single case has to be reviewed by, by a pathologist prior to the data being returned for certain use cases. There's other use cases in research that are highly regulated and a pathologist is the special control on every single case. And then on the clinical side, I mean, our view is, everything we're building has a pathologist as a special control, as a person with the ultimate judgment, as the person who has to ultimately give the final diagnosis and we are.

• 24:31Speaker 01

  In general, aiming to be kind of one component that helps pathologists in the field make the right diagnosis.

• 24:39Jacob Effron?

  Yeah.

• 24:40Speaker 01

  But what the pathologist's judgment is doing will depend on the case. Like, is this a human interpretable output, in which case they can contribute quite a bit? Or is this some sort of phenotype that can only be read by a machine, in which case they're doing more quality control on the overall system? You could imagine, in the same way in clinical pathology or in genetic testing, it's sort of more quality control versus making the primary interpretation because, for certain cases, like reading a nucleotide, a human can't make the primary interpretation. And for some of the features we're measuring for certain, they really can't be read in any other way. So those are some of the factors that help determine the regulatory path. In terms of PathAI specifically, we have 510K clearance for our viewer is one regulatory

• 25:24Speaker 01

  clearance we've had in the past couple of years. And we are working on one other regulated device, particularly in the drug development setting, that we're in advanced stages with.

• 25:36Jacob Effron?

  Yeah. I mean, given that you, on the clinical side, that you use the pathologist's special control, like, what does that, I guess, mean? So every time you have a new algorithm, you know, if you started cancer and then you go into NASH or whatnot, like, do you have to, like, how does the regulatory timeline work around that? And, you know, from, I guess, when you started the business in 2016, like, how long did it take, you know, from a regulatory side to say, okay, you can go, you know, implement this on the clinical side?

• 25:59Speaker 01

  Yeah. So on the clinical side, there's a few different regulatory frameworks. There's the five, like, for example, we have 510K clearance on our platform for primary diagnosis. And then, yes, for each new algorithm product, a decision will be made for how that's going to be used. Is it going to be an FDA-cleared diagnostic device? Or is it more of a research-use-only tool?

• 26:29Speaker 01

  That won't go through FDA regulatory clearance and laboratories are also regulated largely by CAP/CLIA guidelines and laboratories can validate tests under those guidelines. So it's a significant decision for us just based on the time and cost and level of rigor for specific applications for deciding which algorithm products off of our large set we want to submit for regulatory clearance and we certainly do that when required to make the most impactful product.

• 27:07Jacob Effron?

  Totally. I mean, I guess I'm curious, like, taking a step back, obviously, algorithms in pathology just make so much sense. It's something you focus, you know, your majority of your career on. Like, how would you characterize where we are today in terms of, you know, what these models can do well, where there's still a lot of work to be done? You know, what's kind of the state of the field today?

• 27:26Speaker 01

  Yeah, so I think we're still at the beginning of the sort of well-validated, robust sets of products. I think there's been such huge advances just these past couple of years in machine learning that like I've never been more bullish on kind of where the technology is and I'm frequently, my mind is blown by just how well these systems work today or can work. So I think that's like really exciting and kind of, I've thought from the beginning, but it's even more true now. Anything that's visual by eye is going to be able to be done even better with machine learning, augmenting a pathologist than a pathologist alone. So that's a pretty big chunk of pathology and I think if there's two big areas, one is making what we can already see today and the biomarkers we know are important today

• 28:11Speaker 01

  more accurate, more reproducible. And that's actually big and people forget about the importance of both. We focus a lot on accuracy, accuracy is important. Reproducibility and noise is also really important. So just getting rid of noise, getting rid of variability. If you send a case to a pathologist in Wyoming or a pathologist in Boston or California to have confidence you're going to get the same diagnosis every time would be a huge advance for the field. So for that, I think technically we're far and I think a lot of the work today is creating the right workflows that integrate into the clinical practice of the pathology community to really start increasing the utilization of both digital and AI.

• 28:56Speaker 01

  So I think we're still, the field is moving very fast. We're at a stage where the technology is there, we're getting our first high-quality algorithm products, and then, you know, the next big stage will be the integration into practice. And I think for everything in this first category, like AI augmented pathology will be superior to pathology alone. And I always want to make this point because it's so true, they're not either or.

• 29:17Jacob Effron?

  For like everything a pathologist is kind of diagnosing, does it feel like today we have algorithms that are already kind of, you know, could, could help a lot in that diagnosis? Or there are some spaces where it's still been challenging to kind of build those models?

• 29:29Speaker 01

  I think it's been challenging in a sense that people haven't focused on them. So I'd say the answer is no. I think pathologists do such an incredibly large set of different tasks. I don't think it's necessarily for things that are visible by eye a technical barrier. I think it's a, you know, you have to focus on things, for example, that fulfill those three criteria I mentioned. Things where the machine learning augmented pathology is going to be significantly better than pathologists alone, things that are actionable and really address that medical need, and markets that are, are big enough that if you can make this investment, it's going to address a significant number of patients. So I think that's the prioritization, and a lot of things haven't hit that level for the, each of those three. And so I don't think it's technical, I just think it's where people focus, and there

• 30:13Speaker 01

  aren't that many companies in this area, particularly companies who are, you know, building these things really high quality, really robust, and then have platforms that are distributed to actually get them to pathologists. So you do have to prioritize and, you know, look ahead five or 10 years, absolutely, I think the majority. But now I think it's just people, including us, are focusing on what are the most important applications, you know, and then we'll be adding more and more over time. And then the second big area is just what's not being done today but would benefit patients if it were being done. And that's really discovering new biomarkers, discovering new subsets of patients using this technology. And we're at the very early stages there as well. I mean, I see in the future where this technology is being used in every clinical trial. So we're discovering many new important patient subsets who will benefit from specific therapeutic interventions.

• 30:57Speaker 01

  And that sort of use of this technology for discovery and then new biomarkers, new companion diagnostics, that's still quite early.

• 31:04Jacob Effron?

  Yeah, and how does that work on the discovery side? Like, you know, I'm sure it's a combination of these two, but is that process more like hypothesis driven, where you guys are kind of digging into potential areas? Or is it like a total unstructured problem where you're like, let's just look at a bunch of, you know, these slides and clinical outcomes and draw our own conclusions?

• 31:22Speaker 01

  Yeah, I mean, the really cool thing is we can do the full spectrum of that. So with the technology and what's so unique about this approach is you have these inexhaustible images. So there's really no cost in terms of tissue. There's some cost in terms of computer processing, but you know, that's getting cheaper over time and is still extremely relatively cheap compared to the value it could provide. So we absolutely do both. I think one thing that's unique about Path AI is that we cover the full spectrum from fully interpretable, this is the standard biomarker, but now we can measure it more accurately, more reproducibly, more objectively. You know, how good is this biomarker? And that's our baseline. And then we can actually do more, say, end to end type deep learning approaches, where the only, the only sort of knowledge

• 32:06Speaker 01

  we're imposing on the system ahead of time, an extreme example would be just the images and the treatment and the outcome and say, you know, discover. And then we have this whole set of intermediates, which actually often work the best, where we're imposing some knowledge, some prior knowledge, like cells are important, tissue regions are important. But we're saying, we don't know what are the spatial patterns? What are the characteristics of each? Is it nuclear features? Is it others? So let's use deep learning to discover those novel patterns from the data. So to go from a single biomarker that you just measure more accurately, reproducibly, sort of discovery of novel spatial phenotypes, and then kind of the pure end to end approach. It's very powerful to be able to do that full spectrum and then benchmark results. And then you often end up with, you know, the strongest predictive model that you could then say take into a later stage trial or to help design a new trial.

• 32:51Jacob Effron?

  That's fascinating. I mean, I think one thing you mentioned too is obviously, you know, the underlying models and just tech across the industry has gotten better and better. And obviously, we're, we're in like kind of an AI hype cycle right now. And, you know, I'm curious, like, there's obviously been these huge advances in underlying foundation models and kind of just general purpose models, right? I think even in the computer vision space, you had the Meta team with their segment anything model. And it feels like across the board, these, these multi-billion parameter generalist models are getting better and better. Does that influence like your day to day work at all? Or is it still like, you know, the same models you guys were building before? These massive multi-billion parameter models are still the most relevant?

• 33:25Speaker 01

  It absolutely influences our work tremendously today. And I think it just. And I think these are like super real advances and makes us able to be far more efficient with our use of resources and much more productive by, you know, using foundation models where you have these very large models, so fewer models, but more parameters per model, but certainly each model trained on more data and far more generalizable. So the issue in pathology, the area where it is, there's many relatively small applications. Taken together, it's a huge set, it's a huge market, a huge set of applications, but each one is relatively narrow. So if you do have to have a full development process for each one and there's really no synergy between them, it becomes extremely costly, both in

• 34:10Speaker 01

  terms of time and money, to do that. Whereas if you can start with a foundation model and really just do fine tuning as needed with very little additional training for each new application, that's very, very powerful. And that's something we're absolutely doing. It has been leveraging a lot of the recent advances, more generally in computer vision. And you mentioned segment anything model. I mean, things like that, like Facebook showed can dramatically scale and increase the speed at which and reduce the cost of getting supervised annotations for training downstream supervised models. And we're absolutely incorporating advances like that into our process because I do think we're seeing sort of orders of magnitude shift in productivity and decrease in cost based on leveraging some of these new approaches.

• 34:53Jacob Effron?

  Yeah, super interesting. I feel like there's this question of like the extent to which you'll have generalizability across these massive models. And if you're seeing that already in pathology, I mean, that seems like one of the hardest use cases for these models to be able to generalize across. And obviously, I'm sure there's a ton of fine tuning you do on each specific use case, but even just, you know, the model learning, you know, being able to transfer learnings in one place to another is really exciting. I mean, to your point, given the long tail of smaller use cases within pathology, they may not have the clear ROI that like the first few and after due, it's awesome to hear that like, you know, they may require less kind of compute novel work to spin up in the first place.

• 35:29Speaker 01

  For sure. I mean, part of it's because we've, over the years, we're in our seventh year now, have developed a very large corpus of data to enable us to build these foundation models, like Massive sets of whole slide images, massive numbers of annotations from pathologists. We have over 450 pathologists providing us with annotations. We've accessed to over 50 million whole slides that can be imaged. So, yeah, once you have this very large corpus of data, and that's not widely available public data, so once you have a very narrow, vertical, wide corpus of data, yeah, you can build foundation models that will then generalize across quite a large number of tasks within that domain. So I do think that's where things are going.

• 36:12Speaker 01

  I'm not sure these general models build on entirely different images, but I do think in different verticals, those who can accumulate large data sets of enough scale will then be able to build these foundation models that will cover many of the applications within a vertical.

• 36:27Jacob Effron?

  Totally. And I think part of what I think is really interesting about what you're doing, too, is to maybe the beginning of our conversation, you talked about how a lot of pathologist work is relatively qualitative today, and you have folks that have been doing it for a while that can really nail these diagnoses, but maybe it's harder to transfer to newer folks in the space. And I think one thing you guys, I believe, have focused on is this pathomics work of standardizing some of the measures for these diagnoses. I'd love to hear a bit more about that and how you've thought about getting them adopted in the industry at large.

• 36:58Speaker 01

  Yeah, no, so I think it's so important to build off the shoulders of all of the knowledge that has been gained, as a great place where this technology can be valuable. And in many cases, we are measuring the right biomarkers or biomarkers that are very valuable. It's just that they're so difficult to measure based on just how complex the underlying biology is and just the intrinsic challenges of human perception, which is essentially some kind of generative model every time we look at something. We're not exactly seeing exactly what's there. We're doing heuristics to see only small parts of the image because they're so dense, and we're creating different results, different times we see it. And that's why for difficult things, significant levels of both intra within observed variability

• 37:43Speaker 01

  and inter observed variability. So it's not that we're measuring the wrong thing. It's not even that we're inaccurate across a large ...number of people, it's there's just too much variability within within viewer and that there's different biases within different viewers. So some people may just systematically be low and some people are systematically high. So if you average over both, you know, it looks like you're near the middle. But for any individual patient, you know, they may be getting something systematically off based on which pathologist. And even if they show an image to that same pathologist on two separate days, for many very difficult things, they'll get a different answer. So this is a great opportunity for machine learning, which can be trained to be both accurate and is always essentially reproducible in the sense that the same image, a model deployed on the same image will give exactly the same result to the same image. So we focused on this in many areas. I would say

• 38:28Speaker 01

  oncology, some of our work around PDL-1, which is probably the most important tissue-based biomarker for for immuno-oncology, HER2, which is the most important biomarker that's emerged in this new area of antibody-drug conjugates and is critical in a number of tumor types accurately and reproducibly interpreting PDL-1 and HER2. And then another major area for us has been in non-alcoholic steatohepatitis, is making more accurate and reproducible the assessment of inflammation, diseased liver cells called ballooning hepatocytes, amount of fat, and amount of fibrosis. So each of these are somewhat subjective and

• 39:10Speaker 01

  difficult to score reproducibly by manually. So we've trained a machine learning system to aid the pathologist in scoring these more accurately reproducibly. And in terms of adoption, I mean, on the drug development side, we focus on areas where this data is critical for drug development. So for example, the NASH example, the pathology scoring is super critical for rolling trials and for assessing histologic endpoints. So and we're working with the FDA and the EMA to get regulatory qualification for use of this tool in trials. And we think there will be quick adoption because it because it will make pathologists more accurate, more reproducible in the assessment of these biopsies. Similarly, on oncology, we think it's going to be important for adoption to focus on things that pathologists

• 39:55Speaker 01

  are struggling with and that are critically important for treatment decisions. And two examples of that would be PDL-1 and immuno-oncology and HER2 for these new HER2 antibody-drug conjugate therapies.

• 40:06Jacob Effron?

  Yeah, I mean, given all the things we've discussed, I'm curious, as you like, think five, 10 years in the future, like, you know, how does the job of pathologist change? Like, what, you know, what, what, what looks different on a day-to-day basis?

• 40:16Speaker 01

  Yeah, so I think there's going to be a lot of these kind of lower-level tasks, like counting individual cells, or trying to make quantitative assessments of this image data. I think a lot of that work will be done sort of behind the scenes by the AI systems. So you sort of picture now a slide undergoes an automated fashion, an H and E stain, or an immunistic chemistry stain, and then that becomes the raw input that the pathologist has to look at, and then they're essentially doing a very low-level task to interpret, you know, how many immune cells are here, you know, how many cells are expressing this protein brown, how many are negative. And they're, they're doing their best with these sort of heuristics, but they're being provided with very unstructured raw data.

• 40:58Speaker 01

  So I definitely think the future job of the anatomic pathologist is, you're not going to end with the H and E stain and the IHC stain, you will be augmenting the H and E with AI, you'll be augmenting the IHC with AI. The pathologist will be sitting down in front of a monitor that will have a sophisticated AI system that's sort of looked at the patient's medical history, looked at the medical literature, looked at what's going on in the image, and made a set of different recommendations, each with different potential trade-offs. You know, this is what I think the diagnosis is, this is what I think the best treatment recommendation is. And I think the pathologist's job really will be a lot more higher-level judgment, data integration, thinking through these different recommendations, and then really becoming the physician's physician, leveraging all of that to say, you know, this is the

• 41:41Speaker 01

  diagnosis I go for, and really talking through, these are the different treatment, you know, recommendations, to then talk to, say, the treating oncologists, the treating gastroenterologists, who ultimately then will summarize that to help, to help be a good consultant and physician for the patient to make the right treatment decision. I think that will be a big part of the work for some of the most difficult work, and that may actually turn into the majority of the work. I think there'll be certain types of cases in the future. You know, that are very straightforward where the AI is doing most of the work and the pathologist is sort of doing quality control and approving the case. So I think in general, a much larger proportion of their time will be focused on sort of judgment, data integration, and really helping focus on the most difficult cases.

• 42:22Jacob Effron?

  Super interesting. I mean, obviously, you know, we're in the early innings of digital pathology today, but I feel like already, you know, you guys have done a ton on both the research side and the clinical side. You know, I'm curious of all the things you've done, is there like, you know, a piece of work that you're particularly proud of, you know, of the impact you guys have had?

• 42:36Speaker 01

  Yeah, so it's hard to, hard to focus on just one. So I would say in terms of the application area, I think NASH is a really great example of an area where, where there's a huge unmet patient need, there's, it's a huge growing medical problem, there's the unmet drug development need in the sense that it's been very challenging to get effective new therapies approved, partly because of how difficult the pathology part of clinical trials is. And it's something we've been focusing on for about four years now, is building a platform both for new biomarker discovery, but also for more accurate reproducible measures of the key features of NASH. And we're, we really think this will transform the way drug development happens in this area.

• 43:20Speaker 01

  And then that very specific area, I think, really becomes an example for how, you know, this could expand widely across areas of tissue pathology and the way it's used in drug development. So I would say the work in, on our NASH Explore product and our AIM-NASH product are some of the product areas that are, are real highlights.

• 43:36Jacob Effron?

  That's awesome. Well, we always like to wrap up our interviews with kind of a quickfire round where we ask you some, some questions for your takes. And so maybe to start, you know, what are you most excited about in AI and healthcare, completely outside of the work you're doing at Path AI?

• 43:50Speaker 01

  Yeah, so I think there's this huge opportunity for this idea of kind of like personal health assistance that are really with patients a lot, whether they're doing preventative health and they're overall healthy, or even if they have a serious condition in say cardiovascular disease, metabolic disease, psychiatric disease. But there's just this great disconnect between what it's like living with one of these diseases and the way medical care is given today. And it's so suboptimal, where I think this idea of having, you know, large language models and, and intelligent systems that are actually responding to feedback in real time and providing feedback and encouragement and coaching back to people to allow them to more effectively

• 44:34Speaker 01

  prevent disease or treat serious diseases they already have. And that can, I think, incorporate a lot of different things, including sort of compliance with medication, compliance with other recommendations, and make it really empowering for people. Because you know the whole cliche of like, you go to the doctor and they give a serious diagnosis and you like, can't remember a thing they said. And like, the solution today is like, a piece of paper that's very static and has a few things on it. But what if you had kind of like, a real-time partner for dealing with these diseases that are super chronic? So this is entirely outside of pathology, but I think there's tremendous opportunity here.

• 45:05Jacob Effron?

  Totally. I mean, I feel like you're already, you know, I feel like for a while, there was a bunch of these companies that were trying to do automated scribing, like recording visits in doctors' offices and putting that into notes. And the tech didn't really work. And now it actually really seems like it does. And, you know, I think there's, you know, even tremendous potential there to take, kind of, record a doctor's visit and then, you know, these models are very good at kind of summarizing or explaining things at different sophistication levels, right? Whether it's back to, you know, the patient, then to a referring physician, to whoever it is. I think that's a really exciting part of the space to that. And then obviously, you know, you spent the last seven years, you know, getting, you know, AI and healthcare adopted in the system. You know, I'm wondering if you had a magic wand and you could make a change to like, the broader health system or the regulation, you know, what change would you make to kind of speed up adoption of AI in the space?

• 45:48Speaker 01

  Yeah. So I think, I think I'd talk kind of about my, my specific space. I think, I mean, definitely in healthcare, you know, reimbursement and financial incentives are such an important way of driving change. So there's definitely increased costs associated with things like buying digital pathology scanners. There's more and more data coming out showing the improved quality of diagnoses. Diagnoses have a tremendous healthcare impact overall, in terms of, in terms of being the key piece of information for determining down to therapeutic decisions. So I think a reimbursement framework for this technology

• 46:29Speaker 01

  that really recognized the value of increased accuracy, increased reproducibility, and getting patients with serious diseases on the right therapy, that was sort of specifically considering the latest advances in AI, would have a tremendous impact. So I think it would likely be that. And I think more broadly, this sort of move from sort of fee-for-service and service units being generic to

• 46:58Speaker 01

  To really being able to recognize and incentivize use of technology that will increase the quality of care would be tied to that.

• 47:05Jacob Effron?

  Totally. Now, that's a really good point. I mean, I feel like a lot of the way, you know, codes work today is it's around the amount of work it takes for a provider to do. And so actually, when you introduce some of these algorithms and it takes less work, even though it's better for the system and you're getting better results, you may end up with like, you know, lower actual reimbursement for that. And so it feels like, you're right, if you have kind of more of a value-based paradigm, you would be able to avoid that kind of unfortunate reality of the go-to-market there. And then, you know, another thing I wanted to hit is I believe, if I'm correct, that you were, when you were a PhD student, you were in Daphne Kohler's lab?

• 47:38Speaker 01

  Yep, that's right. Yeah.

• 47:39Jacob Effron?

  So one, is it true that all very successful people in health tech just know each other? Because the world seems stupidly small. But two, I'd love to hear a bit about what that was like.

• 47:47Speaker 01

  Yeah, no, what a great group of people to work with. What was it like? It was great. It was one of the, certainly a key to the trajectory that I've kind of been on ever since. So when I went there, I had already done my pathology residency and I kind of really liked machine learning and computer science, but just hadn't done it even in college. So I was kind of like, oh, this is my last chance to really, really dive in completely and learn this new subject. And then, of course, I wanted something that would somehow leverage. I wasn't going to be like the world's greatest programmer or the world's greatest algorithm developer, just because I hadn't done much of it. But I was kind of in a unique position where I could, you know, I have some pathology knowledge, I could link that with knowledge in machine learning

• 48:32Speaker 01

  and hopefully do something useful at the interface. And then, yeah, Daphne was one, she was very generous for taking me on, just because I did have such a different background. Like, she's got the strongest technical, computer science, algorithm, machine learning folks. That definitely wasn't where I was coming from, although I really wanted to get deep in that area. But she saw the value in sort of, you know, the medical applications. And so it was, one, just super generous of her to take me out. I think reflects her sort of, her spirit of wanting to go into something new and really make an impact. And yeah, she was just an incredible mentor and person to train with, both her and the people in her group, in the sense of always wanting to push as far as you can in terms of really doing impactful work, work that's important.

• 49:16Speaker 01

  It's not just, you know, incrementally better than what came before, but it's really a significant advance for the field. It did enable me to learn all of the kind of...

• 49:25Speaker 01

  State of the art, at least learn it as an observer that's having a machine learning and then try to apply some subset of that to the problem I was working on. And yeah, and I think, you know, that became the basis for really, you know, a lot of the same values I've continued to try to apply when I was in academia and into the work at Path AI. So it's definitely a super formative experience.

• 49:46Jacob Effron?

  Yeah, I mean, it certainly, it certainly seems like you're kind of living up to having that impacted more with the, with the work you're doing today. I mean, Andy, this has been a fascinating conversation. I'm sure, you know, folks will want to dig into all sorts of different threads we had here. What's the best way for them to learn more about you and kind of the work you guys are doing at Path AI?

• 50:01Speaker 01

  Yeah, so definitely check out our website. We keep that updated as well as LinkedIn and Twitter. We're, we're always updating both with new advances. If you want to reach out to me, [email protected]. Always happy to meet new people.

• 50:16Jacob Effron?

  Amazing. Well, Andy, thanks so much. This has been a fascinating conversation.

• 50:19Speaker 01

  Great. Thank you. Thanks so much for having me.