Ashish Kamat: Hello, and welcome to UroToday's Bladder Cancer Center of Excellence. I'm Ashish Kamat, Professor of Urologic Oncology and Cancer Research at MD Anderson Cancer Center. And it's a great pleasure to welcome today, Dr. Hikmat Al-Ahmadie from Memorial Sloan Kettering, who is a Chief Pathologist and Associate Attending in the Department of Pathology, and Professor Olivier Elemento, who is the Director of the Englander Institute for Precision Medicine at the Cornell School of Medicine in New York as well. Both of them had a really exciting and interesting session at the recent think tank, which was concluded just about a month ago at which they presented data on artificial intelligence in cancer with specific reference, in some ways to bladder cancer. I'm really pleased that both of you are able to join us today and share your wisdom and insight with the audience. So with that, I'll hand the stage over to you guys.

Olivier Elemento: Well, thank you so much. And Dr. Kamat, it's a pleasure to be here today, really delighted to be able to have a discussion with you and really looking forward to it. So the format of this initial kind of discussion, I think is going to need to involve Dr. Al-Ahmadie and I going over a few slides. We prepared a few slides to sort of focus the discussion and present some exciting applications of AI for cancer, and especially in some cases, for bladder cancer. As you know, there's a tremendous amount of excitement around the use of AI in general, in medicine. And, so what we want to do is to really give examples of applications, have discussions about the strength and limitations, and really hopefully address some of the questions that really have been raised multiple times and are really important to address.

So we'll start with a few slides and a few applications, and then we'll discuss. So really, these slides are meant to illustrate some papers that have come out in the field recently that are really exciting, and they really illustrate how first, in cancer now, we are able to generate a tremendous amount of information as you know about each individual patient. We can do this from biopsies of surgical samples. One of the exciting things, as you know, that's been happening in the past few years now, is the ability to essentially sequence almost the entire genome of tumors using technology such as high frequent sequencing. We can really now sequence the genome from beginning to end and really get a lot of insight into the mechanisms at play in individual patients. And then we are more and more using this information as a way to treat patients.

The challenge is being that as you know, when we sequence somebody's tumor genome, we find a lot of mutations and the challenge is really to interpret those mutations to kind of guide treatment and diagnosis, based on this very complex mutational information that we see in every patient. This is here, an example of how we can use genomic information. This is a paper coming out of Sloan Kettering, a recent paper, that's showing for example, how we can use mutational information from patients. This is data from thousands and thousands of patients. We're showing how we can essentially use information for mutations as a way to predict tumor type. This is incredible because it really demonstrates the connection between the disease and the mutational content of the tumors. And basically shows that if you combine several mutations that you see over and over in patients, you can basically have enough information to say that a patient with a mutational profile is a breast cancer patient or a bladder cancer patient.

And really, one mutation may not be enough, but a combination of mutations adds a tremendous amount of value in indicating what tumor type. There's a lot of interest in using this mutational data, for example, as a way to predict the tumor types for patients whose primary sort of tissue is not always very clear. This slide here is showing the amount, the precision that you get from the mutational profile of patients in terms of being able to predict specifically. For example, whether a patient has non-small cell lung cancer. Very precise, you make essentially very few mistakes when you can quantify mutational content across hundreds of genes and now thousands of genes. And again, the application is multiple, one of them is prediction of the tumor type for cancers, but non-primaries, for example. So this is great again because, this is just an illustration, but this is great because again, it really illustrates the power of genomic sequencing.

And we'll get back to this in the discussion. Now it is again, really a very broad interest in terms of using genomic data as a way to predict response to treatment. And again, this is really the beginning, but there's a lot of work being done. For example, to try to predict response to immune therapy. We know that it takes more than one gene to predict a response to immune therapy. It could be a mutational burden, it could be [inaudible 00:05:35] burden, it could be a combination of express genes and mutations. The future is really towards integrating data from multiple genes, multiple modalities, transcriptomes, and genomics as a way to predict response or diagnosis. So again, this is just an illustration, but we'll get to this more in the discussion.

Hikmat Al-Ahmadie: Thank you again for this opportunity to partner with Dr. Elemento and just to show you some of the interesting and exciting things that are happening in AI. And I shift my focus a little bit into the diagnostics of bladder cancer because there are obviously huge applications for AI and deep learning in the diagnostic realm of care and cancer care. And as the examples that Dr. Elemento showed, basically show you that you really need an important requirement for implementing AI, and for its success is automation and digitization. Without that type of infrastructure, basically, you cannot have AI. And that by itself lends itself really naturally to pathology applications because now it took a while until we were able to get a complete image of a whole slide digitized. Now it's becoming more and more mainstream and it's being exponentially increasing and being utilized in medical centers, small and large and throughout the world. That provides a great opportunity for any type of AI algorithm.

In addition, of course, to the ability to provide diagnostic accuracy and reproducibility which will result in high-level efficiency in assessing these pathologic materials. And not only that, but it will also be very important in the future, or even happening now, to incorporate any potential application for biomarkers, whether that is for diagnostic purposes, for prediction, or prognosis, all of that, are a ripe environment for AI to be implemented.

And something that Dr. Elemento alluded to is that the example that he used was based on genomic, that you can have a high level of accuracy in predicting a site of origin. In this seminal paper from a very strong AI and computational pathology group, the authors actually attempted to do the same, but by using a deep learning and AI approach to determining the site of origin by reviewing whole slide images, and they were able to train and validate algorithms and some of their results are very impressive in achieving a high level of prediction of the primary sites in metastatic tumors when the site of origin was not known. And I lighted one area where bladder was one of the tumors that were assessed and it achieved a good level of accuracy.

So in one study, the authors attempted to reproduce grading of urothelial carcinoma by comparing the automated grade generation with that of three pathologists, who themselves had a moderate agreement amongst them. And then they were able to achieve moderate agreement with the consensus grade that was given by GU pathologists. And another study that employed a deep learning approach to correctly distinguish invasive from noninvasive disease, the authors were able to achieve a very high level of accuracy reaching 96% of the cases.

And similarly in another study, by using a method of semantic segmentation, the authors were able to devise a method and algorithms that could expedite the identification of bladder layers and potentially help in the diagnosis of T1 disease and the extent of their disease. And by applying deep learning approach to urine cytology, the author in one study was able to also risk stratify samples from digital image analysis that was better than a simple histopathology review of the same samples.

So beyond histopathologic evaluation, artificial intelligence has the capability to predict a genetic abnormality starting off the histopathology slides as has been shown now in some publications, by a strong correlation with and detecting FGFR3 mutations and another study by the ability to detect molecular subtypes of bladder cancer from image analysis of the conventional histopathologic slides. And some of these algorithms and applications have been applied to a clinical grade level and have the potential to be applied clinically and enough in the distant future as shown in this study, where the authors were able to show a good level of accuracy in making a diagnosis of cancer that matched the pathologic evaluation, and that can have tremendous help in the sense of allowing the pathologists in the future to focus their effort and attention on cases that are more challenging, that are not straightforward, that cannot be solved by an AI type approach.

So clearly there's a huge potential for AI and deep learning in medicine in general, and specifically in bladder cancer. A few things to keep in mind, that despite the really exciting results that have been reported so far, a lot of them are still a proof of principle, and there's still a lot of validation that is needed. There are potential limitations to AI approaches when you have rare diseases or diseases with not always a straightforward presentations or features. Bladder cancer is a good example and many of its states may not necessarily be straightforward, but I think these things can be definitely overcome. The other thing that is important we'll have with AI, since it's a trainable approach, it's important to have accurate algorithms that will provide the input for the training sets of any algorithm before it can be applied for validation. With this, thank you, and I think we're happy to discuss.

Ashish Kamat: Great. Thank you so much both of you for that excellent summation and summary of the talk and the topic. I think your conclusion slide actually covers a lot of the points that I wanted to raise in my discussion. And I'm going to raise that anyway, so we can discuss that. But the key point there, which I wanted to ask you and feel free to comment either one or both of you, is that there's a perception amongst folks that are not involved in the field, that don't do this all the time, that we could just now have AI take over from pathologists and that's not true, correct? So, what are some of the practical hurdles you see, as far as AI, artificial intelligence, currently, being able to replace the pathologist in day-to-day practice?

Hikmat Al-Ahmadie: So clearly, this is a real fear and it comes up all the time. It may not necessarily be a fear, but it's a worry that what is the AI role, and is it really going to replace pathology? And as we know, pathology, as I mentioned in my slide, lends itself naturally to digitization, and having image recognition is very important and computers can perform very well. And this why, instead of seeing it as actually a competition or a threat, it's the better approach is actually to embrace it and see how it can improve what we do.

There is no doubt that skilled pathologic eyes are very important in identifying diseases. What AI can do if it is trained well, is it can identify a lot of the routine, straightforward cases that may not necessarily need the same level of attention as other cases that may require more human input. So, I think I see a huge advantage of incorporating AI approaches in the diagnostic world because it can save a lot of effort, it can streamline. And we know it is AI is more efficient and is more reproducible than human, no matter how you look at it. So I'd rather use that to our advantage, let the AI help and expedite the routine diagnostic process and leave the more challenging cases or scenarios that would require definite human intelligence and input to be handled by the pathologist.

Ashish Kamat: It's interesting you say that, because one of the places where I think there's a real role for it is in the unknown primary clinics that we have, right? And as Dr. Elemento showed earlier in his JAMA Oncology Publication of the Tumor Type Prediction, we face patients who have metastatic deposits, with no known primary source. Could you postulate, whether you think that a large benefit of an AI, gene-derived tumor type prediction would be in that particular group of patients?

Hikmat Al-Ahmadie: Yeah, definitely. I think that that would be very helpful. And we saw that from the digital slides and the genomics, but maybe I think Dr. Elemento has more to say about other approaches to this.

Olivier Elemento: Yeah. And that's okay. I just wanted to maybe add a couple of comments to what Dr. Al-Ahmadie said earlier which I could not agree more with in terms of the reality, the use of AI as a way to help experts who already, sort of do a great job, but are often very busy. And I think AI has the potential to help in terms of certain tasks. But I think it's really important to remember that. I think we are in the beginnings, in the early days of using AI for medical applications. One of the major shortcomings of AI is really the fact that AI right now, can't really explain itself. There's a major issue in terms of the inability of most AI systems to actually explain why they are actually making a particular decision. And in medicine, that is a critical gap.

The reality is that the pathologist, like Dr. Al-Ahmadie or others, doesn't only just say, "Well, this is the diagnosis." They can also explain precisely why they are making this diagnosis. And that's absolutely critical in terms of the conversation between physicians and conversation between physicians and patients, to be able to explain why the decision and diagnosis were made in the first place. I do think that this is something that the AI field needs to work on. I do think there is potential for improvement. I think that AI cannot explain itself, but it will at some point be able to do so. But this is, I think a major gap, and again, reason why AI is not always being used in the clinic or in pathology as much as it could. It's just this inability to provide interpretation, is a limitation.

Just going back to the second point that you said you were discussing. I do love digital pathology because I think, slides are being digitized and available and there are millions of them that are being digitized. And so there's so much potential to learn from a training slide because we have them and they're so accessible. Obviously, getting genomic information means that we need to work a bit harder. There's just some additional work that needs to be done. But when we have that information, when we can get it and it's just easier and easier to get it, there's so much value in that information. There's value in terms of diagnostic because as we saw, mutational patterns do give you the ability to get a precise diagnosis.

But obviously, there's value in terms of potential matching patient to therapy, right? Because we see that many of these mutations are connected to specific therapies. And obviously, that's also valuable information in bladder cancer. Now, as we know the presence of an FGFR2 mutation means that patients are eligible for FGFR inhibitors. And this is the very beginning of again, being able to match patients to therapies based on mutational content. So there's value in the diagnosis abilities, but also value in terms of matching patients to therapies.

Ashish Kamat: Those are great points, and in some ways, I think the inability of AI to explain itself is a perceived negative, but it's not necessarily a negative, right? Because if AI or pattern recognition or whatever you want to call it, recognizes certain patterns that we just don't understand. There might be something to it, to the tumor biology that we would learn in the future. And from that perspective, do you see that AI will also throw out certain clues for people to start doing mechanistic investigations?

Olivier Elemento: I think that's an excellent point. And kind of like you say, I think this is a point that we don't hear often enough as we kind of learn how to look under the hood, if you want for AI, understand what AI is looking at. I think we're starting to see that AI is indeed able to provide making this sticking information or information that can transform, that can be transformed to make diagnostic information. And that's really, really important. As I said, for example, when AI is making a diagnosis, it's looking at certain patterns, it's looking at certain layouts, certain morphologies of tissues and cells. We need to learn and understand exactly what it's looking at because there has to be valuable information in terms of that where AI is looking and I think we're learning how to do this.

There are some algorithms that are more now prone to interpretation where you can, for example, in a multi-layer, deep learning algorithm, you can look at the weights of sudden connections and that tells you something about where AI is looking at. You can get a map of what we call an attention map, which is essentially kind of a heat map of the pixels that AI is looking at to make a diagnosis. And I think there is really valuable information.

What's a little difficult sometimes is to transform this information into semantic information. AI will tell you that a bunch of pixels is important, but that doesn't always translate into something that we can understand, like let's say a semantics sort of type of information that we can use. And so now I think there is work being done to connect those two worlds, but these are the early days.

Ashish Kamat: No, this is such an exciting topic we could talk about forever, but obviously, in the interest of time, we do have to kind of wrap it up. Let me hand the stage back to you in some ways, and maybe Hikmat, and then Olivier, have you kind of leave our audience with some high-level closing thoughts that you want to share with them on this topic?

Hikmat Al-Ahmadie: Yeah. So yeah, no doubt that AI is already transforming how we think and how we do things, and it's going to be a tremendous help on many levels and pathology specifically speaking from my area of expertise and interest, it still has some challenges to overcome, but I think with time, it's going to add huge value to refining and streamlining our work and will help us improve our accuracy and efficiency. And it will still require a lot of input from us and still needs some input, even at the current stage, especially with the unusual and rare tumor types. But I think there's a huge potential for it and it should be embraced on all levels.

Olivier Elemento: Yeah. And closing thought, I think is really sort of having, really seeing what's happening in the field of bladder cancer. I think there are so many novel therapies that are emerging, being approved. I think these are exciting times now in bladder cancer in the sense that there are so many therapies that seem to be close to being available. I think the next step is really going to be how to combine those therapies, how to match patients to therapies, once you have kind of almost an imbalance of reaches in terms of therapies, you really need to think hard about, who's going to benefit the most from different therapies.

I do think that AI is going to help in that sense. I think being able to predict who's going to benefit from combination therapies, who's going to benefit from a therapy based on combinations of markers, mutational signatures, expression of certain genes, H&E, I think, is really going to be the future of AI and bladder cancer. I think one thing that AI is really good at also is the integration of multiple signals and really being able to use this integration as a way to make predictions that are reliable and robust. I think these are all areas where AI is really prime, really likely to be making an impact in the future, and I think that's really exciting.

Ashish Kamat: Great. Thank you, once again, both of you for taking time from your busy schedule to share your thoughts on this important topic with us. Hopefully, we'll get a chance to actually see each other soon because this digital world, AI, while it makes a lot of sense, I'm getting tired of the Zooming, but stay safe and stay well.

Hikmat Al-Ahmadie: Okay, thank you.

Olivier Elemento: Thanks for having us.