Hikmat Al-Ahmadie: Yes. Hello everyone, and thank you for the opportunity to be part of this discussion. As mentioned, yes, we had, the focus of the session was on AI and the role of basically digital pathology, mainly on improving what we do and what we have been doing in bladder cancer. For that purpose, we assembled a panel of individuals who have great experience in this field. We started with Matthew Hanna, a pathologist at the University of Pittsburgh who talked about bringing AI into the lab, focused on practical and ethical realities of digital and computational pathology. That was followed by Dr. Kong's vision view from the Harvard system and where he talked about building trustworthy AI for pathology and evolving paradigms and adaptive applications. And we concluded the session with the speaker from University of Pittsburgh as well, Dr. Samer Khader, who is a site pathologist and involved in the digital pathology applications in that field. And he talked about digital pathology and AI in [inaudible 00:01:57], its current use and future directions.
Ashish Kamat: Great. Donna, it's always good to have a refresher, not just for myself but for the audience. So maybe you could share with us state-of-the-art AI in bladder cancer.
Donna Hansel: Yeah, of course. And similar to Hikmat's comments, thank you again for allowing me to be part of this. It really was a pleasure being able to work with Hikmat in designing this series that we had at BCAN. It really is where pathology is going, and I was hoping just to take a few minutes to maybe get everyone on the same page who hasn't been following this space. So when we talk about AI in pathology, of course everyone thinks digital pathology whole slide imaging, diagnostics, but we also use it in genomics analysis, using laboratory data and laboratory values to predict patient response and then of course biomarker assessment and prediction of therapeutic response. So I'm going to focus on digital pathology because that's really where I think for a lot of urologists, where we think of when it comes to AI and pathology and a lot of technology changes needed to happen over the years for us to be able to get to where we are today and really enabling AI.
First was development of whole slide imaging systems in the early 2000s. This is where systems could scan an entire glass slide and we could view it remotely. So that evolved early in the 2000s. Shortly after that, Aperio introduced the first commercial whole slide scanner, and this was used for consultations but not approved for primary diagnostic reads. So the field kind of languished in the research space for some time. One of the challenges in the diagnostic space was that to scan all of these slides and Z-stack them, it takes a lot of data, and it really wasn't until 2010 and thereafter, that cloud-based storage evolved to the point that we could actually scan and utilize these images. And then finally in 2017, we had the first FDA approved pathology solution for primary reads. And this enabled more slides to get scanned in, which allowed foundation models to get built for us to be able to do AI.
And so the way, when we think about whole slide imaging, we get a really, hopefully good section of the tissue on a glass slide. This step is pretty critical. It gets scanned into a machine like this and it gets scanned at multiple levels through the glass slide, and that's called Z-stacking. It's basically, as you look up and down through the plane of a tissue, the scanner needs to scan at each of those levels and compilate it into an image. And so as there's integration of these scanned layers, you end up with a whole slide image. And so ultimately what happens is you use AI applications or you develop your own AI algorithm, and what it does is it takes the whole slide image on the far left and it breaks it down into segments, and you can segment into tumor versus stroma. You can segment into inflammatory cells versus tumor, and you can get all the way down to a single pixel.
So what the computer does is really quite amazing and beyond what we could do by eye. And this really has provided a lot of benefits to us in pathology. So there's speed, there's scalability, there's quantification, there's standardization, importantly reproducibility. And this can really assist the pathologist in more rapidly identifying cancer. Theoretically should reduce the workload. We're still in the process of refining some of these tools, and it's just very early tools that are available now. But really the way we're thinking is that this is going to enable precision medicine in the future. So not just histology, but layering on multi-omics platforms into AI algorithms. And indeed there is just recently a paper that came out that looked at real-world deployment of foundation models in pathology for lung cancer. And essentially what they show in this paper on the top in the blue is a standard workflow.
So we get the slides, we cut them, we make a diagnosis, we decide we need additional tests. The clinical team say, okay, if it's this diagnosis, we might need EGFR, we might need ALK, we need to cut it some more for next-gen sequencing. In this paper, what they did on the bottom in green is to take a computational approach and they could digitally identify using AI, whether or not a tumor indeed had abnormalities in EGFR. It saved the tissue. That result came out quickly, within a day. And in fact, if you needed to go on to do more complex testing, you had enough tissue and you could turn that around much quicker. And so their opinion is that you could really get these results with limited physical processing with really rapid output. And what's remarkable when you look at the picture, the closest thing to histology that all of us are aware of was this image.
This is as close as it got where it just showed us how the computer segmented areas of interest. So things are really changing quickly, but it raises a lot of questions, and I'm sure Hitman gets the same questions I do. Does this mean we don't need pathology anymore? The computer's going to make all the diagnosis and who knows? But we're not there yet. And right now we view AI as a tool, not a replacement. Pathologists are still the data stewards. We validate, we interpret. When you layer on multi-omics, it's easy to get spurious results. If the computer is finding something that's not the area of interest, it might be looking at something incorrectly.
There's some famous examples out there where AI goes wrong in identifying the wrong thing, but we do feel that this is bringing forward new opportunities in computational pathology, algorithm training and oversight and really will probably take us to a new level, including imaging three-dimensional tissues, whether in fresh state or formal, in fixed state directly. So I think we're really on the way to a new era in pathology, and I hope this was helpful in bringing everyone up to speed. Thank you for letting me share these slides.
Ashish Kamat: That was phenomenal, Donna. That was a great primer, and thank you for actually sharing that with all of us. Your last slide sort of reminded me in 2002 when the robot was coming into surgery, people are like, oh, the robot's going to take over. You're not going to need surgeons anymore. But it's a tool. Right. It's what we use as experts to help our patients. Along those lines, again, we talk about this at the Think Tank, but also we talk about this at the IBCG retreat where both of you were. We're really trying to help patients globally. Right. So let me first ask you, Donna, and then you Hikmat, how do you think that this high resource intensive process of using machine learning AI might hurt or help people in underserved areas? Donna, you first.
Donna Hansel: Yeah, this is a fantastic question. As you know, I'm heavily involved in our global oncology program here where we travel to many countries and look at the resources, but also where the gaps are. I think that there's a lot of opportunity to help. The basic step is really that the slides need to be cut properly, and I think that's a focus area that can be easily achieved. And once that's the case, then algorithms are available not just for us to look at the slides remotely and make a diagnosis, but for example, in countries or locations that can't bring up molecular testing. In the lung cancer example, there's a handful of alterations that really make that initial first step triage of patient therapeutics. And if we can develop algorithms and other tumor types that really help triage patients to first-line therapy quickly in the absence of a molecular pathology lab, I think it could be really a game changer in improving health. Hikmat?
Hikmat Al-Ahmadie: A very important question, and you can look at it as it could be a problem or it could be an opportunity. The problem of course, in these places where the technology is not always readily available as it is in the Western world perhaps, you can think of it as maybe by investing a little bit of basic technology, you can expand what they can do. For example, in the past we used to send people, actually physical people going to these countries to help them. Now, you may not necessarily need to send a group of people or physicians or technicians to do the work. You can set up as simple as a slide scanner in remote places, and then all the other analysis and interpretation could still happen outside these spaces if the technology does not allow for that. But at least the basic scanning that can allow the slides to be transmitted anywhere in the world to be interpreted and to maximize what you can get from that basic infrastructure. So I think it could be a problem, yes, but it could also be an opportunity.
Ashish Kamat: Yeah, I think it's only going to be an opportunity, right? It's how we use it and how experts like yourselves actually steward this cause. Let me ask you one last question. I mean, we could talk about this forever, but one last question because we see here so much about AI hallucinations, and Donna, you sort of alluded to that a little bit when it comes to these.
When we look at people developing algorithms where it's purely machine learning AI, it's not pathology, they're just doing pattern recognition, which I'm not saying is a bad thing, but it clearly is not biology-based pattern recognition. It's purely machine-based learning pattern recognition. Do you think that this modular specific use case, AI is here to stay, or you think at some point these AI algorithms are going to get smart enough with input from you? Of course that it'll be cross-platform. Right. You just have a slide and maybe not in a year or two years, maybe three years, kind of like ChatGPT or Grok. It's just going to become smart all around and the AI platform will not be disease-specific. I know it's a sort of a looking-in-the-future provocative question, but you guys are the experts. So let me put you first on the spot Hikmat, then we'll give Donna the final answer.
Hikmat Al-Ahmadie: Yeah. So the, I mean, clearly what a bad thing is a bad thing, and it's a real, it's not something imagined. The key thing, as you mentioned, what to do with it. And I think our role should always remain to be supervisors, to help interpret this pattern. So this why, as we've been discussing, AI is not going to replace what we do. It's going to make what we do much better, more efficient, maybe faster. But as we know, our feed is always evolving. There will be new things that were not known before, and there's going to always be a human interpretation to validate the findings, to give a better explanation, to put it in the right biological and medical context.
Ashish Kamat: Donna, final word to you.
Donna Hansel: Yeah, I mean, I think there's always going to need to be a role, not just for the pathologists, but for our partners in urology and oncology, to work with us, to take what's seen as maybe a unique population of tumors, that AI may identify something unusual about a segment, and then take it back to understand what that means. Is there a therapy related to it? Is there a hereditary pattern associated with it, right? There's so much we don't know. Pathology evolves constantly, right? Our classifications come out all the time, and we know more and more, but the computer doesn't know that context. It doesn't know how to say this is something important for a patient. This is something we need to study further. We need to build cohorts. The foundation models are only as good as the slides that go in, but they're not diagnosticians in the sense of identifying what's important for patient care. And that's where all of us working together I think, are really going to push the field into the future.
Ashish Kamat: I think that's a great way to end, working together. And again, on that note, thank you so much for taking the time.
Donna Hansel: Thank you.
Hikmat Al-Ahmadie: Thank you.