Aleks: [00:00:00] What does it mean to bring the science into the clinic? What does it mean to do translational research in the space of digital pathology, in the space of computational pathology? What are the steps? How many players do you need to have? And how close are we from implementing the newest discoveries from nature and other high impact journals into the clinic?
These and many other questions are gonna be answered today by our podcast guest, Dr. Anant Madabushi. Welcome my digital pathology trailblazers. Today my guest is Professor Anant Madabushi. Hi Anant, how are you today?
Anant: Hi, good. Thank you for having me. Having me on aleks.
Aleks: So Anant, you are pretty well known in the digital pathology space, and I wanted to introduce you with one title and it was not possible.
When I was researching you, I had actually traveled saying who you are? Are you a pathologist? Are you Ara radiologist engineer? Which university do you work [00:01:00] with? And just to give, to pre-frame it for my digital pathology trailblazers. I’m gonna read this signature of Anant email that I just got before the recording.
Anant Madabushi PhD, FAIMBE, FIEEE, FNAI. Robert W. Woodruff professor, Wallace H. Coulter department of Biomedical Engineering, radiology and Imaging Sciences. Biomedical Informatics Pathology, Georgia Institute of Technology, and Emory University Research Career Scientist, Atlanta Veterans Administration Medical Center.
Anant, could you clarify? Meaning take some time and introduce yourself to the digital pathology trailblazers.
Anant: Yeah, no, absolutely happy too aleks. Anant Madabushi, I guess I consider myself a babe in the AI and pathology space. I’m a little toddler learning about this wonderful, exciting world, I had the privilege of.
Aleks: Just I need to interrupt little toddler of over 450 publications and [00:02:00] over a hundred patents please continue.
Anant: Yes. That’s why I didn’t say newborn. That’s why I said toddler.
Aleks: Oh, ok. Ok. Understand that explains everything.
Anant: But yes, I’ve been very fortunate, aleks, to have had some really good luck along the way and working with amazing people, and that led me 10 months ago to Atlanta, the beautiful city of Atlanta, where I am a professor in the department of Biomedical Engineering.
And as some of your listeners may be aware, the biomedical engineering department at Georgia Tech, and Emory is a joint department, so they’re two separate universities, Georgia Tech as well as Emory. But they have had, for the last two decades, a shared department, a joint department across the two schools, and was really thrilled earlier this year where the department was ranked as a number one biomedical engineering department in the country.
So that was nice. Little surprise coming into the program last year. But I have also been involved with other departments, [00:03:00] given the nature of my research, working in radiology, working in pathology, and working obviously in biomedical informatics. One of the things that Emory University did for me was to provide me secondary appointments in these areas cause of the nature of the work that I do.
And so even though my primary appointment is in biomedical engineering, I have courtesy or secondary appointments in these other departments. Now I wanna say take a second to talk about the VA because that’s something that I’m very passionate about, something that I take very seriously. And so I had the good fortune about four years ago to get funded through the VA program, the Veterans Affairs Medical Center in the United States, which again, a lot of your reviewers may or may not be aware that the VA enterprise in the United States is one of the largest medical enterprises and is.
Has been developed exclusively to cater to the health and wellness of our veterans, people who have served in active military. And I was really privileged to get a grant from the VA back in 2019 around some of the [00:04:00] work we were doing with AI and lung cancer and that allowed me to formally enter the VA program.
And I was a research health scientist. I was extremely honored when, a few months ago I was elevated to the role of a research career scientist within the VA, and so I know it’s a little complicated, but basically what it means is I have my academic appointment.
Aleks: A little complicated is understatement Anant.
Anant: Complicated, yes. But allows me to have my academic foothold within Emory. But at the same time also have this other external appointment within the VA. And in, in some essence, I essentially have two research groups, one within Emory and then one within the va. But obviously the groups work very closely together and there’s a lot of synergy and a lot of alignment.
And if you are thinking about, as we are in our group about the impact of this work, about the impact of technology for precision medicine, for the betterment of our veterans, you really need to develop these cross connections and Crosstalks, and I’ve been very privileged [00:05:00] through Emory and Emory Healthcare, through Georgia Tech and then the VA to be able to take advantage of this.
Interesting, but like you pointed out, somewhat complicated ecosystem.
Aleks: Not only this, and I’m gonna touch about the later in the episode, you are involved in different industry endeavors and you consult for different industry organizations, which is another complexity level on top of your already complex academic career.
But I’m gonna ask you about this later. Just one thing. The grant with VA, is that the 80 million grant or is the 80 million a different grant that you got awarded?
Anant: So we have several grants. We don’t have anything in the 80 million range. I think the 80 million might be referring to the totality of funding that we’ve managed to accrue over the last decade and a half.
But we’ve been very fortunate to have been awarded a number of grants from the VA most recently, one of the grants that I’m very excited to talk about, most recently, we were awarded a fairly large grant from the VA to focus on prostate cancer, where I think this will be [00:06:00] of interest to your audience.
Where we’re applying AI with pathology, with radiology and bringing both radiology and pathology into play and into synergy with AI to be able to build better predictors of disease, outcome and disease aggressiveness for our veteran patients. So that’s something that’s just kicked off that we’re very excited about.
Aleks: This is great, and I’m gonna keep throwing like those qualifiers because what I have experienced so far with scientists such as you, very accomplished one. I dunno why you guys don’t brag. You just put out papers, like amounts of papers that are unconceivable for me. But I think I need to be an advocate for you to brag about your work.
So expect me to interrupt and tell my listeners what that means, a grant or what that means. Whatever you do whenever I know it, because there’s so much that you did that I have no idea about. But to get to the topic of this episode, I just reviewed the little video that you have in your email signature, and I’m gonna put it in the show notes as well.
And there you said [00:07:00] something that was so simple and profound that your mission or you are driven by identifying the problems that have maximum clinical impact. And this is something that I still see such a huge gap between the academic research and the clinical applications and today I wanna focus on the image biomarkers.
This is, I wouldn’t say it’s a hype anymore because this has been there for probably longer, but I started working in that space in 2016 and tissue image analysis company and finding a tissue biomarker and non-invasive biomarker to stratify patients on images was the holy grail pharmaceutical companies.
My question to you is, how far are we from implementing this into the clinic? And I’m gonna refer to one of the papers that you wrote. And for those who are watching on YouTube, I always have those headphones and they like cover my earrings. But you may see what I have on my earring, it’s a multinucleated [00:08:00] giant.
So I am referring to an old paper it’s from 2021, but we had a conversation on Twitter when I posted a picture of a mouth nucleated giant cell, and I just love them. These are my favorite cells and you can call me quirky pathologists who like finds favorite cells, but I think all of us do that. But the publication that I wanna talk about, and then I let you talk, is the publication from Journal of Clinical Investigation, computerized Tumor Multinucleation Index.
Muni is prognostic in P 16 plus oropharyngeal cancer and this is what happened in this paper. There was something that pathologists were already quantifying the number of those multinucleated giant cells in the tumor tissue, and you took it to the next level with image analysis, applying deep learning and different methods and quantified this better.
And I think in every paper that describes an automated quantification with image analysis, there is this, always this sentence, but pathologist evaluation is subjective and like inconsistent and all that stuff. I think there [00:09:00] should be like a copy paste thing. We could post in all those papers because this is so true and also so obvious.
Anyway, but let’s talk about this type of biomarkers, so those that can be visually quantified or visually confirmed by pathologists. And then let’s talk about the second part of this image biomarker area of research where you cannot quantify, I call them low hanging fruit and high hanging fruit biomarkers.
Let’s talk about the muni example and low hanging fruit.
Anant: Sure. So aleks, I want to touch on something that you talked about before. Forward where you said that you couldn’t quite figure out whether I was a bioengineer or a pathologist or radiologist, which thank you was very flattering, but that’s actually a very astute point that you made.
So first of all, I want to qualify that I am actually a very lowly biomedical engineer, but I’ve had the privilege of working with very smart pathologist, oncologists, radiologists, and that has very much essentially set up my worldview, at least as far as AI [00:10:00] with radiology pathology, precision medicine goes.
What that means is that as a biomedical engineer and as somebody who works closely with clinicians, I have picked up, and perhaps you call it my U S P, if you will, but the fact that I picked up enough of the lingo and I picked up enough of the understanding of pathology to be a little, let’s say, mildly dangerous, and what that means is that I started to really think about these problems from the end user standpoint.
I really started to think about it from a clinician standpoint, a pathologist standpoint, and thinking about the technology, you think about the end user and think about how do you develop a solution that addresses a particular problem in a way that’s going to be satisfying and intuitive for the end user.
And so when we were looking at the problem of head and neck cancers and just a little bit of context about head and neck cancers, P 16 positive or H P V associated head and neck cancers tend to actually be a fairly indolent [00:11:00] class of disease. So this is not to say that all of these cancers are going to do well, but the vast majority of them do well.
Unfortunately, we still don’t really understand or know. Which cancers are going to do well and therefore could benefit from de intensification of radiation therapy versus those patients, the small subset of patients who are gonna do poorly and therefore we need to have the more aggressive regimens for those patients.
And so when I was working on this problem with Jim Lois, a very close friend and amazing head and neck pathologist, Jim and I go back about 14 years now, and we were talking about this problem and he talked about how. He looks at these large aggregates of cells and we spent a lot of time looking at the slides together and he would show me these examples of these large aggregates of cells.
And he said, Anant, I think that there’s something here. And he had published a paper a few years before that showing that just visual assessment of multinucleation was associated with poor outcome. And he said, but this process makes my eyes bleed. It is an extremely difficult, challenging problem.
Aleks: I can still relate to everything that you like, ask a [00:12:00] pathologist to visually assess.
Anant: Right, it’s one of those eye bleed problems and so the other challenge was what you raised, right? This issue of inter reader variability. And if you think about inter reader variability, the challenge goes up significantly when you’ve got a more difficult problem that you’re trying to address.
If there are things. That are easier to visually identify, then you find that the reproducibility is much high, but the tougher, the primitive or the hallmark or the biomarker that you’re trying to identify, so too is the inter reader variability or the lack thereof. And so when he posed this problem, it got me thinking, okay, as a biomedical engineer, this is.
Very appealing because we have the tools, machine learning, deep learning. I think we used a generative adversarial network, a particular kind of deep learning algorithm for this particular work. But that wasn’t the secret sauce. The secret sauce was basically identifying this particular biomarker that had been previously visually identified as having association with poor outcome and now coming up with a [00:13:00] quantitative metric or quantitative index.
That captures the outcome associated with this quantitative metric and also thinking of it from a translational standpoint. Again, I go back to my roots as a biomedical engineer and as a biomedical engineer. You have translation as part of your D N A. I think a lot of biomedical engineers self-select into this space because they’re not only interested in science, they also wanna see the science move forward and actually have impact and have translation.
And so to me, as I think about. A particular biomarker or an AI solution. I’m always thinking, is this going to be reproducible? Is this something we can stand up across multiple sites, multiple centers? And if the answer is no, then even if it potentially could be a very compelling paper, we end up probably ignoring that and going on to some other problem where there is a view or a roadmap that ultimately leads to translation.
Aleks: And just to give.
Anant: Yes.
Aleks: A context for this paper. Thousand 94 patients were selected from six different centers. [00:14:00] So exactly that is to testify that this was one of the things that could be generalizable.
Anant: Exactly Aleks, and thank you for bringing that point up because it’s a very important point. We took a long time to put this together and you can see also one of the things your listeners will no doubt appreciate is a number of co-authors on this publication. A large number of co-authors, large number of sites involved, and that was very intentional because from the get go we wanted to evaluate the validity of this biomarker across multiple sites, multiple labs.
And that’s why we actually had over a thousand patients so that when we validated it, there were no questions about the generalizability. Cause if you want to translate these tools, you have to be able to demonstrate that it works across different labs, different sites and so I was really pleased with that work.
Just to wrap up the question where you asked, okay, how do you translate this and where next? So I’m very pleased to report that we are actually in the process of validating this work in the context of clinical trials and that’s a big.
Aleks: Congratulations.
Anant: Thank you.
Aleks: Cause that was [00:15:00] the, the paper was three years ago. Three years ago, and that was like, oh, we would love to see it in clinical trials, but this is like what every other paper says when they don’t have enough proof.
Anant: That’s right. That’s right.
Aleks: So I’m super, super excited that you guys are actually taking this to the next step.
Anant: That’s right. So it’s been a journey and I think that paper was. Aleks, I will say that if you think about translation is complex. It is very non-linear, right? But overall, it’s becoming clear that ultimately to translate these technologies, you need to be thinking about validation in a few different ways, right? You need to think about validation first from an institutional standpoint, and that’s, I think the J C I paper demonstrated that this biomarker was resilient across different labs.
Second step now is to look at completed clinical trials, which is what we’re doing, and demonstrating that this assay, this biomarker, actually holds up in terms of predicting outcome in the completed clinical trials. But what we’re already starting to do is starting to put this into [00:16:00] prospective clinical trials.
Not from an interventional standpoint just yet, but from an observational standpoint. In other words.
Aleks: Investigational endpoints in trial.
Exactly, thank you. Exploratory
Anant: endpoints, exploratory secondary endpoints. And so to me, it’s becoming clear that if you want to translate these technologies, particularly from a prognostic predictive standpoint, you need the institutional validation to begin with the completed clinical trial validation and then prospective deployment to get it to the point where you understand the characteristics of the assay in a real world setting, and that sets the stage for deployment in a prospective interventional setting.
Aleks: So a question here, since you started this work and now we are in this stage of exploratory endpoints. So not yet the interventional deployment last step, how many years have passed? Just to put it into perspective, how much time it takes to take something that is so scientifically promising and actually [00:17:00] prove it in the real world.
Anant: I will give a more surprising answer. I will give the answer that you don’t expect.
Aleks: Okay?
Anant: If you can do, if you can demonstrate that the data is there, and that’s why we waited so long to publish that work because we wanted multiple groups. And we wanted to publish it in a journal like JCI with its impact factor and all that, I think if you can demonstrate that and get the visibility around your work, it opens up opportunities quite rapidly.
And I will say that from the publication, like you mentioned two years ago, we are validating this in completed clinical trials and we have managed to put it into a prospective clinical trial, an institution, an investigator initiated clinical trial at Emory University, where we’ve already started the prospective validation deployment, which.
If you think about it in a two year timeframe, it is actually not bad. It’s pretty rapid.
Aleks: It’s very fast. Just to give context. I counted in the meantime I counted how many authors are on the paper. It’s around 20 people if I counted well, and [00:18:00] 12 different affiliations.
Anant: Yes, it was a lot of work and creating a lot of. Lot of conversations, lot of data use agreements by the way, which we don’t talk about as we talk about ai. There’s a lot of work that is involved in getting all the institutions on board and sharing the data, but I think that people understood the importance of this work and they banded together and that partnership and those collaborations continue and continue to move towards translation of this biomarker.
Aleks: That brings me back to me telling you to brag more about your work and I see, okay, so you put a lot of effort into having very sound generalizable science. You put it out there and then some doors open and then it goes fast because people already see the value.
They don’t have to question, they don’t have to like second guess. They know the data is there.
Anant: And now that I’m on Aleks’s podcast, things more, even more doors will open up.
Aleks: Of course I’m gonna be now following all your papers and like making little snippets on social media. Hey, look at this, but Anant, [00:19:00] so this is fantastic and I believe it without second guessing, just after reading the paper because the pathologist in me says, oh, I can see those multinucleated giant cells.
I am confident that a computer, once the computer can recognize it, okay, and I can visually verify how good this recognition is, it can count it better than me. I’m fine with that. This is, no-brainer for me. Low hanging bio image biomarker fruit. What about the other type of biomarkers? Where I see the potential of predicting molecular markers that are on images, but without any way to visually checking it.
Not even, with different technologies as attention and whatever you guys are using to like point where the information comes from. The image from the pathologist in me still says, huh, like, How much can I believe it? I know it is confirmed by a different method, but what happened in between?
How can I be sure [00:20:00] that there is no mistranslation of this information from one endpoint to the other? And how confident are we in this other type of biomarker? And how far are we from deployment? Maybe not far at all, but maybe still a lot very far.
Anant: Yeah.
Aleks: Even though I am promoting digital pathology as much as I can, the pathologist in me always thinks really?
Anant: So aleks I’ll give you a slightly long torturous answer here. So I’ve talked about this in many presentations and if you look at some of the YouTube videos, my fondness. Or my preference for more engineered handcrafted features, call it what You will, is quite well known, at least for those people who’ve who followed my work.
And the big reason for that was when we have taken a complete black box based approach, right? Which is an end-to-end deep learning framework. It’s always come up with somewhat surprising results. And I’ve talked about the fact that in 2017 we published a paper in plus [00:21:00] one, where we looked at endo myocardial biopsies to predict cardiac rejection and heart failure.
And we developed a C N to learn from about a hundred biopsies and then, Took that C N and applied it to another separate set of a hundred patients and predicted which patients were at risk for cardiac rejection, which patients were not. And we also had two pathologists look at the same set of images and come up with their predictions of which patients had cardiac failure or not.
And the pathologist came back with an average area under the curve of 74%. The network came back with an A U C of 97%. 23% differential. And so I was at Case Western at the time. I got super excited. You talk about bragging well, great example of where I should have been more circumspect and not gone off and bragged because I went ahead and bragged and I’ve been paying the price for that ever since because I put out a press release.
The press release got picked up. G reports called it one of the coolest things happening on [00:22:00] the planet that week. I was trending on Reddit. Briefly, I was in the top 10. I was on, I was training on Reddit.
Aleks: Congratulations.
Anant: And it was great. Yes. The problem is, Aleks, is that two months later when we got another tranche of images from the same lab, the same institution, and we ran it through the same network, the network went from 97% accuracy down to 75%.
Aleks: Okay.
Anant: And that was.
Aleks: Level leveled down with the pathologist.
Anant: It leveled down and it got to the level of what the pathologist. Were performing it, and even though I’m not too smart myself, thankfully, I’m surrounded by really smart people like my students who were able to find out after about one or two months that what had changed between the first tranch of images and the second tranch of images was a remote software upgrade that had been applied to the slide scanner.
And that remote software upgrade had very subtly changed the appearance of the images that had then resulted in the network. Performance going from that 97% down to 75%. So essentially you had a classic case of, [00:23:00] batch effects and these pre-analytic variations that were impacting the network.
And that was a real important lesson because it made us realize that when you don’t understand how the black box is working or what the features or variables of the black box is picking up there can be a real issue, there can be a real problem. And in my talks, I very often reference the work of Samir Singh, who a few years ago was at uc Irvine, and in 2017 he reported work where he and his student tried to train a network to identify Huskies from wolves.
And it’s very challenging for a human because those animals are very similar and the network came back with a 99% accuracy. And I joke about this all the time, saying that I should be more like Samir, because Samir didn’t go and put out a press release. He thought about why the network could not possibly be that accurate and figured out that the network was speaking up Snow in the background when Huskies were, because Huskies.
Are always present in a colder climate, and that’s what the Husky images, that’s how the images are being recognized. It’s a really important lesson that [00:24:00] I really internalized, and since then our focus has been to leverage the power of deep learning and machine learning, but to do it to identify primitives or biomarkers.
That are more visually interpretable and more instantly recognizable and I think it’s fascinating. I think the work that’s been coming out around virtual staining, being able to predict point mutations like KRAS EGfr P53 and so on, I think it’s very exciting. The concern that I have is, again, this is.
Perhaps just because of my experience is going back to the heart biopsy example. It’s what else do we not know? What else is the network picking up that we don’t appreciate, that we don’t understand? That’s really the big concern for me and so it’s.
Aleks: You should do, or you somebody should do, put up a paper with all the examples where the network was wrong from, so the first time, that was not the first time I heard about batch effect, but something Andrew Janowczyk is always talking about, oh, and the pathologist puts a red dot on the slide.
Network can be [00:25:00] very well trained on that and scanners everything that can cause batch effect. But I think the personal scientific experience with this, Only makes you more aware of that on a different level because in theory, everybody knows that. Everybody knows, oh, different scanner can cause different results.
But then I look at an image from a different scanner and I’m like, yeah probably yes, but I see it. Why should the network be confused? So I think there should be a paper like listing top 10 Deep Learning and Pathology Mistakes that caused somebody So and so much money.
Anant: That’s right. That’s right.
Aleks: And then we could all reference this one just like.
Anant: You’re exactly right. I think I, I don’t think we talk about the mistakes enough, right? I don’t think we talk about, we end up presenting and talking about, this is true not just for me, but several other groups. The problem is we talk about our successes, right?
Because unfortunately, the academic environment we’re in, you’re rewarded for your successes. You’re rewarded for the stuff that worked. You [00:26:00] don’t talk about the stuff that didn’t work or where you got it wrong. And I think we need to talk about this, and this is one of the reasons I’ve been very intentionally talking about our experience with the endo myocardial biopsies, because we got it wrong.
And you gotta own up to that and say, okay, that was a mistake. How can we do better now? The good news is there is a post script to the story, which is that experience. Taught us that we need to be more intentional, and we subsequently published a paper in 2021 in the European Heart Journal where we went instead of looking at 200 biopsies, we actually looked at 2300 biopsies again from multiple sites and showed that when you use the power of machine learning or deep learning to go and identify individual cells like cardiomyocytes, like lymphocytes.
And you create features that look at the architecture and the arrangement and the interplay of cardiomyocytes and lymphocytes. That set of features, that set of spatial statistics actually results in far better prediction of outcome and failure and does it in a consistent, reproducible way. So there was a good postscript that came out of that experience that we started with.
The [00:27:00] black box learned that the black box. Could not always be trusted, but there are ways in which you can use a black box to identify perimeters of interest, but then use more engineered features to then provide your outcomes. So the deep learning is not being thrown out with the bathwater, but it’s still being used to identify the targets of interest, but then more engineered features are being developed once deep learning is done.
Aleks: I see two patterns. So I see this trend of quantifying what’s already being seen with better tools. And I, sometime ago I interviewed Dr from my clinic and I was impressed, like how many classes his model had. It had I don’t know, 13 classes, and I’m gonna link to this episode in the show notes as well, but I was like, oh my goodness.
Complex model, but yet you can verify if it’s okay by a pathologist. So it brings the pathologist back into the picture, into the team, which I’m personally super happy about. The second thing is like unfortunately the magic [00:28:00] pill of deep learning is not that magic. And you have to know where to apply the tool and how to validate and how to verify, which I mean, it’s life.
We should not be surprised, but I think it’s a human nature that you always try to have something that like trumps everything else and you just can focus on that and can skip all the tedious and mundane stuff that you had to do so far. Yeah, a little bit back to the bench, which is good and leverages the potential of the method in a conscious and ethical way without actually like focusing on this.
It’s just by working with the pathologist again, working with the image, with the features that already were suspected by people who look at the images that they can have something to do with prognosis, but obviously visually it’s not possible to quantify. So that’s a super cool of deeplearning.
Anant: Yeah, absolutely. And I will just add to that. One of the things that I also want to qualify is that we shouldn’t be ignoring the potential opportunity for deep [00:29:00] learning to inform and for deep learning to help us with discovery. Since you’ve asked me to brag a little bit, I’m gonna do a little bit of.
Aleks: Go ahead.
Anant: For a paper that we just published two hours ago. Actually I just.
Aleks: LinkedIn.
Anant: Yes. I just posted it. And so it’s a great example where we use the power of deep learning to focus on immune cell clusters or immune cell niches in the context of lung cancers. Also very large study. We looked at almost 1800 patients across multiple sites, both in the United States as well as in Europe.
And one of the breakthroughs that we made in this paper was that with the power of deep learning, we were able to identify sub clusters of tumor infiltrating lymphocytes. There’s been a lot of work on using machine learning and deep learning to quantify tails, quantify the number of tails, and then correlate with outcome.
One of the things that we discovered in our analysis with deep learning here was that we were able to identify subsets of the tools and that some of these subsets of tails looked a little different compared to the others. Now.
Aleks: Based on h and e?
Anant: This was just of h and e, just of h and E, right? And [00:30:00] what we were able to find was that some of these. Clusters of tills on the h and e’s were more strongly associated with outcome or response. And there were other clusters that were actually not really doing very much that they were not really telling us about outcome. They were not telling us about response, and we did a fairly elaborate study. I’m not gonna bore you with all the details, but we did a number of things in this paper.
Aleks: I’m gonna link to this paper as well.
Anant: Sure, sure.
Aleks: Everything that we’re mentioning is gonna be in the show notes for those who wanna learn and read about it in more detail.\
Anant: So one of the things that we did in this study was we identified these sub clusters of tale. And found that when we looked at adenocarcinomas, that when we coregistered the h and e images with quantitative immunofluorescence on a cell by cell basis, we were able to identify what the subtypes of the immune cells within those clusters that we had identified on the h and e were and we did this for the adenocarcinomas.
We did it for the squamous cell carcinomas as well. And then we also did a gene set enrichment analysis off the Klux [00:31:00] clusters. So we were able to actually find associations with biological pathways. Why I’m telling you the story is that it actually shows you that deep learning could be powerful because it further enhances our understanding of the immune milieu that exists within this very complex tumor microenvironment.
While we have classically on an h e image, just looked at all the tills as one monolithic entity. Our data is showing that there are clusters that we could identify directly on an h and e image. Now, of course, on a quantitative immunofluorescence image, or if you look at spatial transcriptomics, we know that you’ve got CD four cells and CD eight positive cells.
You’ve got a T-cell repertoire and a B-cell repertoire. We know that, but to be able to appreciate that directly from the h and e images is very exciting, and it’s also telling us something fundamentally important about a lot of work that’s happening. With till quantification from h e images, namely that if you’re trying to come up with ways of prognosticating outcome or predicting therapeutic response, maybe we shouldn’t be [00:32:00] quantifying all the tills because there are tills that are activated.
There are tills that are really contributing to the outcome and the treatment response. And then there are maybe your bystander tills or your exhausted cells, or you are stressed. Immune cells, which are not contributing. And I just wanna put in a word here for deep learning because it provided a way of making some discoveries and insights that we may not have been otherwise able to come up with.
And so I think that the bigger message is that we need to think carefully and deeply about how we’re using these approaches and making it sure all times that we’re trying to connect it back to biology. And so in this study we did, like I said, a very elaborate correlation study with quantitative immunofluorescence, with gene site enrichment analysis and really focused on understanding what that signature was.
But a lot of papers that I’ve seen, unfortunately, don’t do that. It’s a lot just end to end. And then also some retrospective, intuitive analysis of what the network might be picking up, where you generate visual [00:33:00] attention maps and so on. And I think those are useful, but I think we need to push ourselves as a community to try to go further.
Now I understand the resistance. The resistance of course, is that it means that we need to understand pathology better, we need to understand the domain better, you need to understand cancer better, but isn’t that part of the job? Isn’t that what we’re supposed to be doing.
Aleks: That is part of the job and how I see what you guys just did in this paper, and I’m gonna read it in detail, maybe make a little video on it as well. You guys deconstructed the black box outside of the black box. So when I worked for the tissue image analysis company, this and still is immune oncology research is based on ihc biomarker quantification, on immunofluorescence, on multiplexing and doing exactly what you just mentioned about teals and different immune cells in different entities too.
Like classifying them and characterizing them. So far, there was no way to extrapolate it to the h and e in a way that’s understandable by the researcher, but there was this process of taking h [00:34:00] and e, doing end to end, and having the final outcome. So how I see what you guys did, you basically deconstructed what’s inside of this concept, this end-to-end concept on a local basis, it can be verified by a tissue X or by a pathologist, by tissue researchers. So that is super cool.
Anant: Exactly. And Aleks, I want to link it back to something that you talked about in the beginning, which is on industry and industry collaborations, and Yes, let’s talk about that.
And that’s something that is part of the reason I mentioned this paper because this collaboration, this partnership did not just involve other academic sites, but it actually involved our pharmaceutical partners at Bristol, my Squibb, where we worked with BMS over the course of two years, worked with them to get access to one of their clinical trials, check made oh five seven, and then validated the signature in terms of predicting clinically relevant outcomes.
In patients with lung cancer who’d been treated with nivolumab in this trial from bms. And so it is really critical because the, I go back to [00:35:00] demonstrating the validity of what we are doing. And the J C I paper that you referenced earlier was great, but didn’t actually have the clinical trial validation.
And so when we were developing this work, we were intentional. We said we wanna get a lot of data and show that this works on a lot of patients, but we also wanna make sure that we are validating this. On completed clinical trials, and that’s where our partners. BMS were very helpful by providing access to the clinical trial dataset that allowed us to provide a higher level of evidence that this signature is real, and it is showing an association with clinically relevant outcomes.
Aleks: This is amazing, and this just proves that you had enough proof of concept for a major pharmaceutical company to evaluate and immediately buy into that and give you access to this data. This data is super protected. It’s not that, anybody can just go and ask for this. Not only is this ip, but this is like huge, I don’t know how to call it milestone, huge value data for future [00:36:00] therapeutics.
And I love how you are the personification of what you said, that you’re working on something that’s gonna be brought into the clinic and you, like personally and your team are doing it. It’s not like putting all those publications into the ether of PubMed or wherever and let somebody who’s looking for it find it.
You’re actually like taking it with you. You’re having enough proof for major pharma players. You have the resources, have the, that’s what they’re doing. They’re bringing those drugs. To the patients. So is the next logical step to have this as something incorporated or any of this incorporated into clinical practices, or is that not the next step?
Do we stop by at the drug development or is there an arm that actually goes to the practices where pathologists can have. The software that quantifies it or however, interact with these discoveries that you guys are doing because from the logistical standpoint, there is nothing preventing it. You like run your algorithm on a whole side image or [00:37:00] not even, you could take a picture and run it and have it right.
There at the point of diagnosis. What are your thoughts on that? How far are we?
Anant: Yeah, it’s a great question Aleks and I can give a little bit of context, a little bit of history also about the translational journey, which I, wanna make sure. Also, I recognize that sometimes when scientists like myself, investigators put out these stories and we talk about, we’re working with pharma bms, it feels oh wow, you know this is great.
These. These guys have some secret sauce. I just want to talk about what it’s taken, right? So this journey to get to where we are right now has been a journey that’s about, been about 20 years in the making, right? It didn’t happen overnight. So for the young scientists, students, or postdocs, or thinking about the faculty career, I wanna say that it is, Highly rewarding, but it is not gonna be easy.
And it is a non-linear path. A non-linear journey and it started for me about 15 years ago when I started my first couple of companies. I was very passionate about translation of the [00:38:00] technologies into impacting healthcare and patients in a positive way. And I started a couple of companies, made some mistakes.
Along the way, one of the companies is focused in the cardiovascular space. Another company was in the breast cancer space, and this was back in 2008, 2009, the company that was in the breast cancer space, this was in 2009. I was talking about AI or machine learning before it was ai, about how we could predict which patients had worse outcome with breast cancer and which patients had better outcome and we could do as well as molecular tests.
And I remember people telling me, almost laughing at my face. In fact, I remember very clearly somebody in 2011 or 2012 saying to me, looking at me with a straight face and saying, so you are telling me that your machine learning can identify features on these h and e images? That can tell us which patients are going to respond to chemotherapy or not.
And I said, yes. He looked at me and said, I’ve got a philosophical problem with that. That was something I never forgot and I said, you know what? I’m gonna make sure that [00:39:00] I just prove this person wrong and prove to myself that there is value here. That is tremendous information that’s later in these images, and that with machine learning, with image analysis, we can unlock the information in there.
So it wasn’t easy, I was trying to get funding for my first company, the breast cancer work. It was a challenge. Ultimately, the company got acquired by another entity. And one of the things that I realized in the translational journey was that we, as scientists love the research. We love the papers, but when it comes to the next step of translating, We end up shying away from it because it’s messy and it’s somehow unclean, right?
It’s, it’s business and business is dirty and, we should think about science and research, which is pure. And I just realized that was, for me personally, it was not I was not being true to myself or true to what I was really passionate about, which was I realized that what I really wanted to do was not just to do the science, but to get the science to patients.
And that was a journey, it was a [00:40:00] journey which involved doing the science publishing in the high impact journals, using that as a basis to open up collaborations, get access, like you mentioned, to clinical trial data sets that I might never have otherwise been able to get access to build up the primary data, filled up levels of evidence that then allowed us.
To take it to the next step, to translate it and to deploy it. Ultimately, what ended up happening about three years ago was I ended up starting another company and with that company.
Aleks: What’s your company?
Anant: It’s called Picture Health. I wasn’t sure whether it was okay to talk about it.
Aleks: Yeah, totally. Any you can, whatever you can mention. Mention.
Anant: Okay.
Aleks: Brag about your companies, research everything. It’s a digital pathology business podcast and science.
Anant: Yes it’s called Picture Health. We’re very excited about it. Thrushane Rule is the c e o of the company and a number of my former students are part of the company and the company is essentially realizing the promise of the data and the publications that we generated in the lab.
So essentially the picture health has now become the [00:41:00] commercial vehicle or the commercial engine. Take these discoveries forward to be able to translate it. It’s building off the collaborations that we established with our pharmaceutical partners in the lab, and that’s provided the basis for the company now to move forward with these existing partnerships and existing relationships and move these technologies forward.
So it’s exciting. We still have some way to go, but I just wanted to put it out there for your listeners, particularly some of the ones who are earlier in their career, that it’s a long, convoluted nonlinear process and you gotta keep at it. But the important thing that I realized is if you are passionate about moving your technology into the workflow and for it to get out there, don’t expect somebody else to do it.
Nobody else is going to do it. You have to take ownership of that. And that is challenging because as a scientist, you wanna focus on the science and it’s very comfortable to focus on the science. You don’t wanna go talk to investors, you don’t wanna have to, learn about a convertible note.
Aleks: You want your science to speak for you, and everybody’s hoping that it’s just gonna be so good that you don’t have to go and [00:42:00] market it.
And I need to insert here something that I see, like you again, you’re the personification of something that’s happening in the industry. So your steps of the journey are current silos of the industry, you have the scientists that, like I say, guys brag more. They don’t wanna brag, they wanna do the science.
You have on the other end, the commercial partners who are looking for some cool technology and they wanna start selling it. There is a lot in the middle. Like you said, the science is regarded as so much pure than the commercial side. There is no viability with the commercial side, the labs, and the places where the diagnosis occurs.
Where the patients are being treated, they buy stuff. If it’s not on sale, if it’s not on the market, they cannot take advantage of this technology and they have to buy stuff that is validated, that is solid, that they can be sure that they can use. They’re not there to investigate and perform three year long validations.
They’re there to take something that has been proven. To improve patient care [00:43:00] and start using it. So it’s like a whole chain that currently is siloed. I like to believe that my role with this podcast and with whatever I’m doing on social media is to bring everybody in the digital pathology space together to translate.
You said that you like learned enough pathology to be dangerous. I like to think that I learned enough of the machine learning and AI concept. And software development concept to be little dangerous as well. And to start challenging people we’re like, why not? Why not? Why not? And like we’re at the both extremes of pathology and computer science.
There’s a lot of people in between. And bringing them all together in the digital pathology space is my mission. And I see you doing it extremely well at scale as just, you have a team, but you are the team leader. So really tremendous respect.
Anant: Thank you.
Aleks: And congratulations on every, everything. And also congratulations on pulling in the people that you already know from your science world that are going to provide this quality to the market of pulling them in, into your startups, into whatever [00:44:00] business activities you’re doing.
Anant: So Aleks, I’ll just say quickly, one of the, one of the things that, that the greatest compliment that I get after giving a talk is to be mistaken for a pathologist or to be mistaken for radiologist. So it’s the greatest compliment because it’s an important point that you touched on, which I think your listeners should particularly the ones earlier in their career should really appreciate is mult linguality to be multilingual, to be able to comfortably talk about T-cell repertoires or B-cell repertoires to be able to comfortably talk about an a plasia and a plastic cells.
To be comfortably able to talk about adenomas, to talk enough to understand the pathology enough that you are able to have a serious conversation with your clinical colleagues. One of the things that I’ve very intentionally tried to do and I try to communicate this to my students, is understand the literature enough to not just know stuff, but start to bring that into the way I’m talking to pathologist.
I’m talking about pathologic complete response. I’m [00:45:00] talking about new adjuvant treatments. I’ve taken the effort to understand that I read about clinical trials, I understand what the clinical trial is doing, what is the context in which a clinical trial is being performed, and that requires additional effort and it requires you to get out of your comfort zone.
But it is that multilinguality that I think ultimately will drive the best science. In my humble opinion.
Aleks: Now that you mentioned this, I’m gonna brag about something that I have because this is a common gap that I see, and I saw it when I first started in this space. I was the pathologist. I had to learn the computer science speech, and I was working with computer scientists.
They could not speak pathology either. I did develop a course, it’s called Pathology 101 for tissue image analysis, specifically for tissue image analysis scientists and people working with algorithms to learn. What they are seeing in the tissue, what kind of immune cells are there? How to talk to a pathologist with a specific module, pathologist glossary, and couple of different things, and I’m gonna link to this in the description as well, but this gap that you’re addressing, like how did you learn it on the job, [00:46:00] right?
It took you, how much time did it take you?
Anant: It’s been an ongoing, continuous experience. I’m continuously learning, constantly learning, and I, one thing that I just realized is the importance of listening when you’re having these conversations. Listen a lot and buy bit, internalize it. I’m very grateful and I think we talked about the aesthetic talk that I gave.
One of the things that I talk about in the TEDx talk.
Aleks: Yes. We talked about it before we started recording. Yes. So let me give a preframe to that. Anant is at TEDx talk speaker officially. As of, I don’t know, when you recorded last week when we had to cancel the recording because you were busy with TEDx talk and it’s gonna be out in a couple of weeks.
Anant: That’s right. That’s right. Yeah.
Aleks: Tell us about the TEDx talk.
Anant: Yeah, so the TEDx talk really chronicles my journey from India. When I was a graduate student, I lost my aunt to breast cancer. That was one of my first inflection points that really promoted me, intern or pushed me forward in terms of, Developing AI machine learning to address cancer, to address oncology.
And my second major inflection point, [00:47:00] which I talk about was meeting Michael Feldman. Michael Feldman, who was a very close friend. Michael and I met at the University of Pennsylvania. He was an assistant professor in pathology.
Aleks: I’m in Pennsylvania, not at the university, but based in Pennsylvania currently.
Anant: Yeah. And he was a junior faculty then, and I met him and we gotta know each other. He introduced me to judicial pathology. He introduced me to thinking about AI applications for whole side images. In fact, one of the first papers on ai, machine learning for visual pathology images came out from our group back in 2005 or 2006, and that was another major inflection point because had it not been for Michael, I would probably not have gotten into the space.
Michael is now the chair of pathology at Indiana University, and so I talk about some, how some of these personal professional inflection points really dictated the career and dictated the mindset and that initial intersection with Michael just brought home to me the importance of understanding the domain that you were working with and that’s why I’ve been very [00:48:00] intentional and all the collaborations ever since where I’ve worked closely with pathologists, I still try to listen as much as I can because.
Yeah, the technology is great machine learning, ai, deep learning, but really you need to understand the problem you’re solving, right? And I realize that the more you listen, the more you internalize, you understand truly the problem that is gonna be solved, and then you can think about the solution.
I think too often, a lot of our eager computer scientists rush to the solution without really understanding the problem first, and then you come up with a solution that is not very useful to anyone. And so understanding the context of who your end user is. Understanding that linguality, like I said, that mul those multiple different languages, truly understanding the problem is just so critical.
Aleks: Thank you so much, Anant. I feel like we could spend two more hours talking about different things. I might note like what I wanna ask you, but I hope I can steal you for another episode, for now. Thank you so much.
Anant: Sure.
Aleks: I’m gonna link to all the resources that we mentioned to your company page, to the papers in the show notes, so [00:49:00] everybody who wants to have a look at that, it’s gonna be there and thank you so much for joining me.
Anant: Thank you. Thank you for having me, Aleks. Cheer.
Aleks: Thank you so much for staying till the end. I know this was a long episode, but at the same time, I feel like we only scratch the surface. I mentioned in this episode that I have the course that are gonna help you speak pathology, speak pathologists, and understand what’s happening in the tissue.
This is a course for everybody who’s starting their career in the tissue image analysis space. So I’m linking to this course below. Be sure to check it and that’s gonna be your short. To confidently navigating pathology knowledge necessary for tissue image analysis. So be sure to check that out and I talk to you in the next episode.
