Transcript 0:00 When will AI meaningfully change the trajectory of cancer diagnostics? There are very few AI tools in this space which has got FDA approval. How do we standardize the quality and the diagnosis of the data? 0:12 Do you think we will have Spotify for precision medicine? Absolutely. [upbeat music] Hello, everyone, and welcome to the Inference Show by Turing Post. 0:27 And today I'm speaking with, um, Akash Purwarikar, an AI scientist and product leader working at the frontier of digital pathology and computational biology. 0:38 He leads the pathology map platform at HistoWiz, focusing turning complex histopathology data into insights that accelerate cancer diagnostics and precision medicine. 0:48 This is a very exciting episode for me because we decided to air it and dedicate it to Breast Cancer Awareness Month. So thank you for joining me, and let's, let's talk about this very important topic. Amazing. 1:00 First of all, thank you so much, Ksenia, for inviting me to your podcast, and I'm super excited to be here and, uh, just discuss some of the, like, pressing things and what's happening in the field of cancer diagnosis. 1:13 Why is breast cancer so highly misdiagnosed, especially in its early stages, and how is AI sort of helping bridge the gap? That's actually one of my first questions. Let's start with the big question. Yeah. Big picture. 1:24 When will AI meaningfully change the trajectory of cancer diagnostics? What I believe is that, like right now, we are in the stage of digitizing the pathology labs. 1:35 And as you know, even now, the clinicians diagnose cancer under a microscope. 1:40 But now there's a large push towards digitizing those large volumes of glass slides, and that is what will sort of open the doors for bringing in like intelligent tools like AI to sort of like, you know, better detect the cancer features. 1:54 So right now we are at the transition phase of g- going digital from physical, but then I think we are running on two parallel roads. 2:02 Whereas one road is like, you know, towards digitization, and the second road building AI tools that will help assist the clinicians in a positive way to better diagnose. 2:11 So what I believe is that, like very soon, maybe in the next couple of years, it will converge to a point when, wherein, you know, you have all your glass slides digitized and you have your cutting-edge AI diagnosing features which might have taken a lot of time or which might have been misdiagnosed early on. 2:27 If you think in terms of patient outcomes, what is early detection, better precision, what breakthroughs are real right now and what's still aspirational? Right. 2:36 So I think right now where the field is to get the adoption from the clinicians who are the domain experts for cancer diagnosis. 2:42 So the AI is being built for the existing literature that, you know, doctors have been using for quite some time. And there are like, you know, some features which takes a lot of time. 2:52 There are some features which are highly missed. 2:54 So right now, uh, AI is being built as a diagnostic assistant tool to like, you know, just like speed up the process and, uh, also detect like the features which they do like every single day on a routine basis. 3:07 But then we are also seeing some moonshot projects wherein, you know, there are some AI based on like LLMs and deep learning, wherein it's being built, uh, to look at the features which previously the doctors might not have paid at- attention to, but it like, you know, it's a very important prognostic marker. 3:24 So but those are all like long, long moonshot projects. 3:27 But right now to get the quick adoption, let's keep it simple and, you know, try to build AI models that, you know, the doctors can use it in their day-to-day practice. 3:36 And when you say that the doctor's using it for quite some time, how long they have been using AI that you observed, and what's changed from the initial, you know, what's this and how they think about it now? Right. 3:50 There are like hospital systems in the US, like the Mayo Clinic, Mount Sinai, NYU, MSKCC, wherein early on they were, uh, using AI on the digital images to look at high-level features like tumor, necrosis, stroma, and all that. 4:03 Because of the maturity of the AI algorithms in the space, they are now using it for low-level features. 4:09 It could be like cell detections, types of cells, and not just that, the spatial heterogeneity among these features. 4:15 So they are trying to ask much more complicated features, complicated questions, even beyond feature analysis. Like, can the AI sort of map out the understanding the spatial tumor heterogeneity within the tissue sample? 4:27 And, uh, what prognostic markers can I get from doing this? So I think we have taken a few steps ahead to what, you know, the way they were using it previously and right now. 4:36 But as I said, the features that had, they have, that some of the hospitals have been using in the past, we want to push it to more hospitals at a nominal cost in the US and outside. So. 4:47 You are an AI scientist, so you know the machinery- Yeah... of AI pathology, how it works. So AI pathology isn't one model predicting cancer or not. 4:59 It's slides and pictures and research, the volume of diagnosis, multimodal inputs. Yeah. So what is the hardest about working with whole slide images? Yeah. 5:11 Can you, can you like really go deep into the machinery, how AI works there- Sure... specifically? Yeah, absolutely. So I think that's a, that's a very important question that you asked. 5:21 So when we are dealing with whole slide image, one image is anywhere between a few hundred megabytes to a few gigabytes. 5:27 As I've said in the past as well, that is if you take, uh, one image printout at a forty X resolution, it occupies anywhere between one to two tennis court. 5:37 So that is the volume of image that you're looking at, which is very different from a cat versus dog or like scene detection, which is much, at a much, uh, lower resolution. 5:45 And the doctors typically detect at a forty X resolution, so the magnification is quite high. 5:51 So again, this comes with a problem that, you know, we are trying to find like a needle in a haystack problem from a computer vision perspective. 5:58 But also when you talk about like memory constraints and all that, now we want to load this entire multi-gigabyte image or like multi-hundred megabytes of image onto a server and run AI at scale. 6:11 So there are like a lot of like parallel processing that has to go in place to run your AI model. 6:15 But typically in the machinery, what we do is we do like tile patch-based processing, where we chop the entire image into multiple patches. We run the AI on the entire patch images, and then we group it back. 6:28 So again, this is a very like, you know, uh, standard like routine that, uh, we AI scientists follow. 6:34 And, uh, differently, I think in the field of pathology and I think outside the field as well, it's very important to understand the data. 6:41 Yes, I might be a great computer vision scientist, and I might have built amazing models for other types of data, but when it comes to pathology, it's very important to understand the features that you want to look at and how does it differentiate from one category to the next. 6:53 Some sort of like a literature review or sitting with a pathologist is Crucial because at the end of the day, you're building the AI for the pathologist, and, uh, to some extent, it should be done by the pathologist too. 7:06 Not in terms of coding, but you should have gotten like, you know, a significant expertise and communication from early on when you're building the AI. 7:13 And I'm more than happy to dive into detail, like, you know, if, if you want me to. Yeah, that's interesting. 7:18 Uh, because the way you explain, it's, it's more on the computer vision side, but then there is also generative AI now and like- Yeah... how that plays out. Right. 7:27 I think, again, the adoption of generative AI for diagnosis or diagnostic, like, decision assistance might be like, you know, I would say it's too low right now because, uh, we are in the stage wherein, you know, the AI is not able to tell its failure modes. 7:46 Like, it doesn't know, like, you know, when will I get a false positive or a false negative, especially when you're dealing with, like, these large LLM models and all that. 7:53 Now, if you take out, like, you know, the few short learning approach where you train it on a smaller dataset, well-curated, captures the diversity of the sample, there is a possibility that, you know, once you build an AI on that, it can actually tell you, like, you know, where the AI does good, where, where it does bad, the kind of data, the nature of data and all that. 8:11 And that was the core of my research during my PhD where I, I built an explainable AI framework on a subset of data, but it could actually tell not just like, you know, where it performs well, but where it fails. 8:22 I think again, it- it's taking baby steps towards adoption. You cannot come up with a model which is gen AI based and try to entirely replace the pathologist, and they are not going to buy it. 8:31 Neither will the patient be have an acceptance to, you know, be diagnosed by such an AI agent, as you would call. But, uh, it has to occur like linearly, at least for now. So we are in that linear phase of, like, growth. 8:42 You said the word agent. Would you say that agentic workflows are implemented or have been implementing in the workflow of practitioners, or it's still too far for that? 8:54 It's still too far for, like, pathology, but I think there are multiple pieces within the pathology workflow. Like, one thing is you do the biopsy, you send the slide for processing, you digitize the slide. 9:04 Next is you have a workflow for quality control. You wanna make sure that, you know, the image that has been digitized is of highest quality without any pen marks, blurry regions and all that. 9:14 Now, you can incorporate an agentic workflow to, like, you know, make sure the quality is standardized across the entire cohort and is adaptable or amenable for the AI that has to come at the next stage. 9:25 So I think, uh, you can take parts of the workflow and create agentic AI models for that. 9:30 And, uh, I think we are talking about, like within HistoViz, the company that I work for, we are talking about creating agentic workflows for standardizing the quality of pathology data. 9:39 But I think, uh, we have to envision this as more like chunks and not the entire piece, like for diagnosis. So that's way too far out. 9:47 So from what I heard, what the bottlenecks are, explainability and, uh, let's call it hallucinations or false replies. Is that all? What are other bottlenecks that you see, and when will we deal with them? 10:02 So I think the biggest bottleneck that I see even during my PhD and research is, uh, the lack of availability of high quality annotated dataset, especially in the field of pathology. Why? 10:13 One, I think, you know, the patient has to be consented to make it available for research, and we're talking about pathology images. Could be breast, colon, liver, any organ. 10:22 But not just that, there is a high degree of disagreement among the domain experts while creating those annotations. 10:29 Even in my PhD, when we got the annotations for the breast sample, for the pre-invasive, we took the consensus diagnosis or, you know, majority voting among the three expert pathologists who were like, you know, generating the diagnosis or annotation. 10:42 So you can imagine that, you know, your ground truth data might itself be imperfect, wherein, you know, there might not be like a gold standard ground truth for that particular image. 10:52 There could be a doctor from NYU giving it a particular diagnosis, but if you take the same sample to West Coast, they might diagnose it differently. 11:00 If you take it or the same, uh, sample to Europe or India, it could be entirely different. So I think that is the biggest bottleneck, and, uh, that is what the field is struggling right now. 11:09 Like how do we standardize the quality and the diagnosis of the data? And not just that, like, you know, the biases on the data itself. 11:16 So it's, it's heavily dependent on the hospital system whom you collaborated with in training your AI model. 11:22 But once you train it, how adaptable is it to, like, patients outside US or outside ho- like, you know, outside your hospital or even outside the region? So that's the biggest bottleneck. 11:33 Do you mean it can be that specific? Yeah. So the thing is, uh, if the AI model was trained from the data from, say, NYU, right? 11:40 Now, NYU has its own, like, you know, training for the pathologist, where they, like, you know, come together and they look at some features for diagnosis, but it's not gold standard for all over the world. 11:51 If you take the same image to, say, Stanford, they might diagnose it differently. They might give some other category. 11:57 Now, your AI might completely fail because their ground truth might be entirely different for that particular sample, and we are talking about pre-invasive. That is where the maximum confusion lies. 12:06 So I think that that's the biggest bottleneck in, in the field of pathology right now. That's very tricky. So you building PathologyMap, is that something that you address with this platform? 12:17 So the platform that we're building right now at HistoViz, it's an image management platform wherein, uh, it will allow you to, like, you know, manage, access and analyze your high quality images. 12:26 But, uh, in addition to that, we have also created an AI orchestration platform that allows you to ingest AI apps from multiple companies, and the user can test. 12:36 For example, they can run, uh, say, a breast analysis from four different companies on the same slide at the same time. That can tell you, like, how the AI is performing and, you know, the intricacies and all that. 12:47 We want to give the decision-making to the hands of the user in what AI tools they want to run. 12:52 So that's why we didn't make it exclusive to one or two AI apps, but opened it up and made it non-exclusive, open, so that, like, you know, we can ingest multiple AI apps or even upload your own annotations that you might have generated using Python, MATLAB, or R onto the platform. 13:07 And, uh, we are one of the first ones in the world to do that. Is it more for medical specialists or for everyone? It's mostly for, uh, preclinical and clinical pathology research. 13:16 By preclinical, I mean it's for the drug discovery companies for clinical trials, where they want to understand how the drug works on human data, on mice data, and they can work on that. 13:26 Or clinical data, wherein, you know, they might have a patient, uh, five years ago ran a, like a breast biopsy. 13:32 They want to upload it to a digital platform because accessing the physical slides might be cumbersome and might take a lot of time. 13:38 Now they have a digital platform wherein, you know, it's easy to access, easy to share, and they can also compare their clinical notes. 13:44 So that's what we're trying to do with our orchestration platform called PathologyMap. What about regulations? How is that treating you? Because such a- Yeah... sensitive topic. Absolutely. 13:54 So at Histoviz, uh, we are GLP compliant, so we are in preclinical research. So we don't run the AI for diagnostic patients, but mostly for, like, you know, historical patients or animal models. 14:05 So for us, the regulations are slightly relaxed. But for the field itself, I think there are very few AI tools in the space which has got FDA approval. 14:13 But if you look at it, and I'm more than happy to attach, like, you know, a few resources to the comment section or description, the FDA approval has been obtained for a very specific task. 14:23 It's not the AI for, say, prostate diagnosis, but, you know, observing, like, high import- like, you know, high important regions within the image. It's very specific. 14:32 And, uh, uh, still the industry is struggling to get FDA approval for the apps that actually makes the most sense in the diagnosis setting. 14:39 Let's go back to, uh, what we started with about Cancer Awareness Month, Breast Cancer Awareness Month, because I want to ground this technical side and the human side. 14:48 Um, have you seen moments where AI has already made difference for a patient or a pathologist? Well, I think, uh, yes, to an extent, but all of it is in pilot phase right now. 15:00 And I think to give you a sense, the field that I'm in, which is digital pathology, it came into being only in 2017 when US FDA approved using digital scanners for primary diagnosis. So it's a very, very new field. 15:11 Before that, radiology has all been digitized, and they have been using AI tools for that. But pathology, it's at its, like, you know, inception stage. 15:19 And COVID actually sort of, like, you know, made the field go really fast. It gave it the right push because now the, the doctors didn't want to come to hospital for diagnosis, so more and more hospitals are digitizing. 15:30 The field has evolved in, uh, some in terms of, again, digitizing. But when it comes to diagnosis, I think real-world cases, we are still a little far from that, but still it's all in the pilot phase. 15:42 It- it's not in, like, you know, a real-world setting because, again, you want to be so sure that the AI is doing the right thing. 15:47 And, uh, if you deployed it into a hospital, like you as a patient, will you go to a hospital where there's AI in place, or will you go to a hospital where there's no AI? 15:55 So there's a lot of questions to be asked, like, you know, even for the hospital to bring an AI product into, into market. But having said that, AI is making strands in the workflow optimization. 16:06 As you might already know, US is facing a large shortage of pathologists, and in the coming decade, the gap will widen even more. 16:13 So AI is definitely helping the current pathologists to just, you know, improvise their workflow. It could be quality. 16:20 It could be these features which takes hours together for them to diagnose, like few low-level features like mitosis, lymphocyte counts and all that. 16:28 AI is doing a very good job to identify these features and quantify that. So they're using that parts of their, like, you know, existing knowledge to come with a diagnosis together. 16:36 So in that way, yes, we are seeing a good change. Uh, we should have probably started with that. 16:41 But still, can you give me a one-sentence description what digital pathology is so everyone just understands what it is exactly? Yeah, absolutely. 16:49 So digital pathology is all about digitizing your pathology glass slides, which was always looked upon, like, you know, something that you will see under a microscope for primary diagnosis. 17:02 Now, digital pathology utilizes these high throughput image scanners that scans your physical glass slides, puts it on a big screen, very high resolution, and it allows, you know, integrating your image management tools, softwares and all that. 17:16 You can compare it to the radiology, which was, like, you know, it has been digitized, like, a decade ago. 17:22 And, uh, similarly, like, you know, the pathology labs right now, the routine workflow is that you do a biopsy from the patient, you put it on a physical glass slide, you mount it, send it to the histology processing, and then it sits, it goes to a pathologist office. 17:37 Mm-hmm. And they have a microscope in their office. They look at it under a microscope, do the diagnosis. 17:42 But just imagine having one step forward wherein you have a scanner or a camera that sort of digitizes the physical glass slides and, uh, you know, now the doctors are able to look at those high-resolution images. 17:55 And only you can imagine the powers of digitization. Like, one thing is it's very easy to access. Second thing is very easy to share. 18:03 For example, the patient could be in rural parts of US or rural parts of, like, say, India, but still an expert pathologist sitting in New York or California can diagnose that patient. 18:13 There are, like, very high, large advantages to, like, going digital. I'm just highlighting few of them, but, uh, these are the reasons why, like, it's becoming so popular right now. 18:21 Can you give me your timeline when we will see the main milestones, breakthroughs? I think to take a step back, Mayo Clinic has digitized 12 million glass slides. NYU is digitizing all of its slides. 18:34 That's the first step, right? So again, to draw parallels, you can think of it as before Spotify existed, we all had cassettes and DVDs to watch movies, to listen to songs and all that, but now who uses them? 18:45 It's all digital, right? Even with physical books, like Google, when they launched the Google Books, they actually parsed out the entire book library and made it all digital. 18:56 So there was a huge initiative to make it all digital, and then they built tools on top of that, like there was Kindle and, and all that. So right now we're in the stage of digitizing the pathology slides. 19:06 I would say, like, in the next, like, two to three years, we will be seeing more and more FDA approvals for diagnosis or maybe not just that, like diagnostic assistant, where there will still be a human expert who is the pathologist. 19:18 They will be using all these intelligent tools just so that, you know, they don't have any bottlenecks which, or roadblocks which they might have had previously. 19:24 But I think in the next two to three years, we will definitely see either more cases being diagnosed by the pathologist, more access to high-quality pathologists all over the world through this, and not just that, like within a single case, the accuracies will also increase because, you know, now they all have all these tools to avoid accident. 19:42 Do you think we will have Spotify for precision medicine? Absolutely. That's the goal. Yeah. Do you know if anyone building it? No, but, uh, there are some interesting, like, you know, uh, initiatives by Tempus AI. 19:53 You might have heard of them. 19:54 There is, there's Tempus who are, like, you know, building, like, multimodal, like, data sets and intelligent tools wherein, you know, they are taking a pathology report or radiology report, blood samples and genomic analysis all together and, uh, how well can you, like, you know, drive the precision medicine by profiling the entire patient across multimodal data. 20:12 So Tempus is doing a fantastic job in that space. Well, if we extend your timeline a little bit and, um, think about some ideal situation, how you see it, what will be the human role in this workflow with AI? 20:25 Human role will be the same that we have seen right now. Like, for example, I think of this as your, your Tesla. If you would ask, like, "Where would be the human role when you design the Tesla?" 20:35 it will still be the same. You are still in the driver's seat. We are still very far from, like, you know, going away from the driver's seat and letting the car drive by itself. 20:43 What I think is that, like, it'll still exactly be the same where it was previously, except that you will have- Tools that will, like, avoid accidents. 20:52 It will give you, like, you know, it could be like computer vision tools or AI tools that will help you detect features which you might have missed, and it might help you access, like, previous reports from, like, tens of years. 21:03 Like, you know, you might run the AI on, like, the entire 10-year repository of slides, and you want to compare your clinical notes. Okay, there was this patient seven years back. I diagnosed this patient to be this. 21:13 Can I quickly take a look at the report? 21:15 If you want to go physical, it will take a hell lot of time, and it's unimaginable to go into the, like, you know, the physical building, look at the glass slide, access the report, and all that. 21:25 So I think we will definitely see a very large improvement in how you sort of like collate the data and use it for better diagnosis, but still the human part will still be the same but with much more tools at hand. 21:36 Do you know the economics behind it? Is it, uh, supported by the numbers, the development of this area? Yeah. That's a, that's a fantastic question. 21:44 So I think we're all talking about the ROIs and, you know, patient, uh, reimbursements and all that, so there is an active conversation in this sense. 21:51 But, uh, I think in the long term, there is a mandate in US that, you know, you have to hold the physical glass slides for, I think, close to 25 years. 21:59 So just imagine if there are, like, 12 million patients, 12 million glass slides across 25 years that has to be stored within the hospital setting. 22:09 You would pay much, much more for the real estate cost versus for a digital cloud storage, and you can still hold it in places like Iron Mountain and all that, like, you know, where they can hold the slides. 22:18 So I think storage in the long run will, is much more optimal when you go digital. That's one advantage. And, uh, you know, it's also the case sign-outs. 22:26 So say a pathologist might be signing out, like, um, I don't know, maybe, like, 30 cases with physical. 22:31 With d- digital, they might be doing 50, and that will sort of compensate the shortage of pathologists and, you know, more number of patients who are coming to get diagnosed will, can get it done. 22:40 So, but still, I think this is a continuous conversations with the, the healthcare organizations in US as well. Yeah, that makes sense. 22:47 Let me ask you, what are you concerned about, and what excites you the most about this world that you help build with AI? Uh, for pathology, you mean? For pathology. 22:58 For pathology, I think, uh, I can answer the first que- like, second question, like what excites me the most is, uh, again, for me, it's telepathology. 23:05 Because, uh, for me, there might be patients who might not be part of the Tier 1 cities where, uh, the core, like, you know, specialization or specialty doctors might exist. 23:16 This field allows them to get, like, you know, good care by digitizing the slide, like, you know, remotely sending it to the expert pathologist. 23:23 I think there is this communication platform that happens even for second opinions, right? So there might be a doctor in, in New York, but they have a friend in California. They want to quickly get a second opinion. 23:33 With physical, it might takes a lot of time, but with d- digital, it can happen really quickly within minutes. 23:38 So I think that excites me the most, wherein you're trying to establish a communication platform for the pathology data and the experts through the domain, so that's very exciting. 23:48 And what scares me is the AI developers who don't have domain knowledge and are building AI models for the space and trying to go for the post hoc interpretabilities like the Grad-Cams and hotspots and all that. 24:02 It's truly not explainable. Like, you know, if the AI model doesn't talk the language of the pathologist, you don't have an explainable AI framework. 24:09 Sure, it'll give you papers or maybe patent, but it will not get a clinical, like, you know, approval. 24:14 So that scares me because I have seen several models where they, you know, they just show some heat maps, and there is Grad-Cam and all of that and call it as an explainable AI model, but it's not. 24:23 Unless you can tell me what features are being detected and are those features actually pathologist-friendly. 24:29 You can't develop an AI model and, you know, give some, like, deep learning features, and when you take it to the doctor, they have no idea what this feature is. It has to start small. 24:37 You have to talk the language of the expert. That's a very important point. Yeah. Thank you. My last question will be about books. Is there a book that shape your, your approach to AI and medicine or life? 24:49 And it can be a book that you read when you were a kid or a recent book. 24:53 Since we, I think we are talking about, uh, the pathology and being more technical here, one book is, uh, The Blue Book for Pathology, which is the World Health Organization classification of tumors. 25:04 It's a very, like, understated book. It's called the WHO Classification of Tumours, and they have one for every organ, and there are multiple volumes for every organ. 25:13 So for breast, there are, I think, like, close to 12 to 15 volumes wherein, you know, in each volume, they tell in detail as to these are the guidelines that has to be followed in diagnosing the breast tissue. 25:24 So I followed that throughout my PhD, and, uh, especially if anybody wants to get into, like, a pathology field, it could be any organ. I think, uh, you, like, go and read that and try to understand the data really well. 25:37 That would be one book that, you know, that has, like, stuck with me in my early PhD days, and, uh, I'm sure it will be useful for others as well, so. Thank you so much, Akash. 25:45 That was, uh, deeply technical and also deeply human because you talk a lot about how actual doctors and pathologists work with the technology. Thank you so much, Ksenia. It was very, very nice talking to you. 25:56 [upbeat music]