Intro: Welcome to Precision Pathology Podcast, featuring interviews with authors and global thought leaders who are at the forefront of technological advances in precision diagnostics and therapeutics. Your host is Dr. George Netto, the Simon Flexner Professor and Chair of Pathology and the Laboratory Medicine at the University of Pennsylvania's Perelman School of Medicine in Philadelphia. Here is Dr. Netto.

George Netto, MD (Host): Welcome to the inaugural episode of my new series, Precision Pathology Podcast. We're very excited about this and most excited about my new guest today, Dr. Chad Vanderbilt, Assistant Attending at the Memorial Sloan Kettering Cancer Center. Chad, will be discussing his team's exciting study on developing a novel AI-based computational pathology tool that was just published actually July, 2025 in Nature Medicine.

Chad, of course, is representing a huge team, an impressive lineup of investigators. Many friends at Memorial Sloan Kettering and at the Icahn Mount Sinai School of Medicine, including his co-senior author, Dr. Thomas Fuchs, and the first author, Gabrielle Campanella. Thank you, Chad, for accepting to be the first guest and for accepting the invitation.

Chad Vanderbilt, MD: Thank you for inviting me. It's super exciting to participate in this podcast.

Host: Well, listen, with a catchy name like EAGLE that you picked for the tool. EGFR-- What is it? EGFR genomic lung cancer evaluation. With that and beside more importantly how potentially practice-changing this tool if you really deploy it for all of us, it's going to really impact how we work as pathologists and as oncologists on lung cancer management. So, I figured this definitely should be the number one episode.

So, let's dig in. And as we usually start, maybe you can give the audience some background why you did this study, this amazingly-- I'm sure it's a lot of work, a lot of hours, a lot of months. So, tell me why.

Chad Vanderbilt, MD: Yes. This study is many years in the making. I think we first started working on the challenge of predicting EGFR mutations in lung cancer back in 2018, shortly after I finished fellowship. There was a study from NYU showing that you could predict mutations from cancer H&E histology alone.

And we knew that with older methods, you could only get performance to plateau at some level. And with the advent of the pathology foundation models, it really pushes performance up further. And in this study, we even go beyond just a traditional foundation model. We introduced this concept of fine tuning, adapting the foundation model, specifically specific to a task, and that greatly increases performance to the point where we think that the technology is actually ready to use in clinical practice.

Host: Wonderful. So, probably the majority of the audience are not as expert and versed in the terminology like you are. So, to help, having read the study and enjoyed it. So, you build first, there is a foundation model that you built on what? Over 5,000? That's where you trained initially that model, 5,000 lung cancer cases, adenocarcinoma. And then, the fine tuning is that internal validation dataset and then the external validation dataset, right? Do I understand it correctly? That's how you do it?

Chad Vanderbilt, MD: So, the initial model that we used is actually GigaPath, which was developed by Microsoft. So, GigaPath is built in this self-supervised learning framework where the model is exposed to millions and, in some many cases, billions of pathology images and the model learns features that are specific to H&E histology. What we did in this study is we took that pre-built model built by Microsoft. And it doesn't only work for GigaPath, it can be used with any other of the pathology foundation models that have been built. That includes UNI, which is developed by Faisal Mahmood at Harvard. There's also some commercial foundation models from Paige AI that are built with self-supervised learning.

So, we could have used any number of models as a starting point, but what we added to it was that you expose the model to an end task, in this case, predicting EGFR mutations in lung cancer. And you actually pass the information learned for that task all the way back to the foundation model. So, you're actually updating the foundation model. So, it's a whole new foundation model that is even better at predicting the end task than the original foundation model.

Host: So, that's that training piece. Then, that's where you use the 5,000, the initial images. And then, you went ahead and fine tuned it even more after that on the internal validation, which is a set from your institution and then external to make sure it's generalizable.

Chad Vanderbilt, MD: Exactly. We trained on the 5,000 slides to update the weights. And then, we had an intermediate set of cases that allowed us to understand what the performance looks like in a real-world setting, but it's retrospectively. And so then, we could understand what cutoffs, what thresholds to use to get the performance we needed for the clinical task. And then, finally, we did this silent trial, which is perspective. And so, the models had been trained, we had predetermined what the thresholds were. And then, we ran the model in the background silently, not impacting patient care at the time, but seeing how the model would perform in a real clinical scenario.

Host: We will get to that. That's probably one of the most exciting parts of the design. So, we're already in design. And the silent trial that you're talking about, I want to make sure I understood and the audience do as well. So basically, after you found what's the scores, how the model work after you fine tune it, and pretrial and all that. Because the current way workflow is you get the slides, as pathologists, you digitize it or not, depending on your setting. You diagnose it, you see lung cancer, you may need to use immuno and what have you first to decide if it's adeno if needed, usually a small amount. And then, you order the molecular.

And since NGS, like MSK-IMPACT that in your institution, which one of the best performance, it takes a couple of weeks at least, a lot of us are developing this rapid test, right? And the rapid test, it's a couple of days, takes a couple of days, and it's not perfect. So, you're trying to see can you decrease the number of rapid tests? The idea of the rapid test that you don't want to use all the tissue and wait two weeks because it's so critical to start the treatment first line therapy, if it's EGFR positive.

So, you're trying to assist that rapid test, triage, is there a way for us to just tell by image? Now, insert this AI-assisted model, the EAGLE and say, "Okay, this one is going to be negative for sure. This one positive for sure. This one, maybe. Let's do the rapid test." So, you decrease the number of rapid tests, and you did that, pretending that you would've acted upon it. And that's a silent trial, right?

Chad Vanderbilt, MD: Exactly.

Host: And then, you went back and measured how much savings are there in terms of decreasing the number of rapid tests, turnaround time and all that, while maintaining that it's really the AI model work. Does that describe it? So, silent, really, that's the difference between silent and interventional trial where you were going to treat basically. In this case, you didn't treat based on the AI.

Chad Vanderbilt, MD: That's right. Yup.

Host: All right. Good. Good. Good. So, go ahead.

Chad Vanderbilt, MD: For the rapid tests, we value them because it provides information more quickly. There's just a number of limitations of them, and that is that they're PCR-based tests, they have a limited target range, so not all mutations can be detected by them. They also consume tissue, and we're already dealing with very small biopsies from lung cancer. So if you're using tissue for the rapid test, that can result in limitations in your next generation sequencing testing because you've already used the material for the rapid test. So if we can reduce the number of rapid tests, then you have less challenges in getting the final NGS result.

Host: Perfect. And which is it's important to emphasize that even if you are going to use the AI, you still need a confirmation by NGS later, especially for that group that you're not a hundred percent. So, that's not going to change, but the idea will be you have this ultrarapid method and save tissues, especially if you're somebody like me who's not an expert day-to-day on lung cancer, may order additional immunos just to waste some tissue. So, we don't want that. We want to do everything hopefully by AI. Wonderful. How would you summarize the findings? What's struck you and what were you surprised to find anything beside what you were hoping for?

Chad Vanderbilt, MD: So, the main results that we're most excited about are just the overall model performance. So, being able to predict EGFR mutations, not only on the data, the institutional data that we trained on, but also working with our partners in Germany and in Sweden and at Mount Sinai. Our performance of the model does not degrade when going from MSK to another institution. So, that's a major achievement by our team. And the end clinical task, the results we show are that we can actually reduce the number of rapid tests without degrading the overall performance of our screening technology. So, we can skip the rapid tests on up to 42% of cases and not degrade our positive predictive value and negative predictive value for EGFR mutations. And we could go actually one step further than that from the results of our silent trial. You could actually with high confidence rule out that there's EGFR mutation in about 20% of cases. In which case, if properly validated and there's regulatory approval, you could say, "This slide is definitively EGFR negative," and proceed the clinical treatment that would follow that.

Host: Correct. Because then, you're no longer EGFR, TKI is indicated as first line that you're switching them to chemo and PDL, ICI, immune checkpoint. Wonderful. But that's clearly going to need to await some prospective trial, I guess, or getting first the FDA for this and, wow, how amazing.

So, the rapid tests are things like Idylla, like other tests at different places, PCR based. So, hopefully, this will have the advantage that it's cheaper, clearly, unless you guys are trying to cash out too fast. But hopefully, it'll be cheaper. And more importantly, you know, a point that we need to emphasize because we have some global audience, I hope for this podcast too, it's in low to middle income countries. Part of the problem with NGS is it's not available in a lot of countries, and there are tens of thousands of lung cancer patients who don't have the access to this. So hopefully, this is-- you call it democratization of the technology. Can you expand on that?

Chad Vanderbilt, MD: Oh yeah, absolutely. That's one of the main central motivations of developing AI tools for predicting mutations. The cost to run the model is just the cost of compute, which is essentially electricity. The model can run on consumer grade GPUs. You don't have to have an enormous high-performance computational setup to run the model.

So, it's possible to take the model, which we actually made the weights publicly available under a specific license for people to perform research. So, people can take the model and run them on their slides in their specific setting to test performance, validate the thresholds.

So yeah, this basically allows people who work in places that don't have access to NGS to get very accurate results for the end target, in this case, EGFR in lung cancer.

Host: And I know you have some plans to go beyond EGFR, not just EGFR, right?

Chad Vanderbilt, MD: Yeah, absolutely. It's also just validating the general technique for predicting mutations from H&E alone. So, we have a number of other targets. We have models that will predict all of the common driver mutations in lung adenocarcinoma in a single model and with similar performance to what we see in EGFR. So, we have that coming forward. We also have an interesting development that you can actually take an image with a microscope-mounted camera or even a cell phone from the objective of the microscope, take a picture of the tissue and run them. And I wrote a special pipeline for running the model on an image from a smartphone camera. And the performance is actually nearly identical to the performance we see from the whole slide image scan slide. So, we see this as a great opportunity for democratization.

Host: That's even, yeah, a step further. And so, they just send you the image and you can run it for them and help patients and global health and all that. A couple of things I wanted before we wrap up. It's an amazingly impressive, but I noticed that it works better on primary versus mets. So, I want you to say a couple of words about that. And also, there are some false positive and false negatives. And they're biologically and morphologically for us as pathologists very intriguing. Can you comment on that, why mets don't work as well?

Chad Vanderbilt, MD: Yeah. I mean, part of the challenge with metastasis is that we have fewer examples of it. A lot of the samples that we receive for sequencing are the primary samples. So, there's just a larger number to train on. But we also see that there's an architecture of the lung itself that supports morphology that is specific to the mutation we believe. So when the tumor metastasized to the brain or the bone, it doesn't have the same scaffolding on which it was originally developing. So, we think that that's likely the reason why metastasis are a challenge. But the more work we do and the more data we have to train on, we think we can overcome the limitation of work on metastasis. And it's not that the performance on metastasis is terrible, it's not as good

Host: Yeah. To me, it's that knowledge of maybe these were many of them and the metastasis, sometimes, not always, already have received treatment, clonal evolution, you know, molecular revolution and all these things can play a role. And you mentioned, where are those false positive--

Chad Vanderbilt, MD: Yeah. The false positive is actually very interesting because we actually saw many of our false positives were in mutations that are biologically very similar to EGFR exon or kinase domain mutations. So, HER2 exon 20 insertions are biologically very similar to exon 20 insertions in EGFR. So, we basically see that the morphology overlap is extremely tight as well as exon 14 skipping mutations and MET. So, almost all of our false positives were one of those two mutations. To overcome that, training this multi-label fashion where we provide not just EGFR status, but also MET, ALK, HER2, RET, ROS, we provide all those labels and train individual probabilities of positivity in all of them. That allows it to favor one over the other. So, you might have a high probability in EGFR, but you also have a high probability in the other. And that allows us to understand better, whether it's a potential false positive for EGFR, which is why we state that it still needs to be verified by preferably a next generation sequencing technology.

Host: Amazing. It just goes to say how morphology is the ultimate phenotype of the genotype, so things that converge around the same morphologic features. A great pathologist like you with their eyes, but even AI cannot sometimes will count both as same category similar to each other.

Chad Vanderbilt, MD: Yeah. And this finding validates that what the model is looking at is biological. It's not just finding something that correlates with the EGFR mutation. It's likely the actual biology of the cancer that's being interpreted by the AI model.

Host: Very reassuring, very impressive. And I love it when you said right after fellowship in 2018. So, you've been in the field only really seven years, and look what you brought us, and with Tom Fuchs and others is amazing. So really, the future is that. That's why I wanted to name the series Precision Pathology.

I think as a field we're really moving forward to things like the ones you're discovering, and poor people like me who trained on the scope are going to run out of jobs. But we're very happy to see that as long as they're more precise and cheaper and more democratized, like you said. Thank you, Chad. I'm proud of you and of the team, and thank you very much for coming and allowing me to discuss this.

Chad Vanderbilt, MD: Yeah, of course. And of course, the work that I do, we stand on the shoulders of many folks, the pioneers and computational pathology, Thomas Fuchs, as you mentioned, but also the infrastructure that was buit at MSK under the direction of David Klimstra and funded through the Warren Alpert Foundation, the Warren Alpert Center for Digital and Computational Pathology is key for our ability to do these types of projects, and also the collaborations with Mount Sinai and other institutions around the world.

Host: Yeah. And I have to put a plug for my friends, Meera Hameed and others, because we got all the good molecular ground truths there. So, wonderful. Absolutely, it takes a village. Thank you very much again.

Outro: We would like to extend our sincere gratitude to today's guests for their valuable contributions and thoughtful perspectives. We also wish to thank our dedicated production team and the technical team at DoctorPodcasting. Special thanks to Jennifer Vepler for her dedication and skill in coordinating our episodes.

Please note the opinions expressed by our speakers are their own and do not necessarily represent the Perelman School of Medicine or the University of Pennsylvania. Thank you for listening, and we look forward to welcoming you back for future episodes.