Ashish Kamat: Hello everybody, and welcome to UroToday's Bladder Cancer Center of Excellence. I'm Ashish Kamat, urologic oncologist in Houston, and it's a pleasure to welcome to the UroToday forum two stalwarts in bladder cancer, Professor Zlotta from Canada and Professor Al-Ahmadie from New York Memorial Sloan Kettering.

So both of you participated in this debate at the AUA. This was part of the IBCG AUA Forum. It was really something that kicked off a lot of discussion, both during the debate and even afterwards about the influx of AI into our field. And of course, we've seen AI in multiple other arenas, but when we see AI tools such as this coming into clinical use, coming into use on clinical data and pathology, there's a sense as to are we overreacting or are we under reacting in some ways? So it's great to have both of you come here, sort of relive that debate in a bit, but it's more of a discussion today. So with that, Alex, I'll have you get started first and then Hikmat, we'll hear your viewpoint.

Alexandre Zlotta: Thanks Ashish, and good morning. Yeah, as you know, I mean, artificial intelligence is clearly part of our daily life. You just want to have a look at the next restaurants or the meaning of anything, you type it and AI will tell you what it is. So the real question, what's its exact place in what we do today, which is in NMIBC?

And so there's so many unmet needs where AI, as you know, could help. It may minimize pathological grading. So not that we're going to get rid of pathologists, but at least minimize variability. We wish that all surgeons do the same job. We know it's not true. But both in pathology and for surgeons, I really don't think that AI is up there yet.

There's one place where AI, I believe, can actually already be used by physicians and patients, and it's for prognostic tools. And if you just think, we had that debate. I mean, we recently published a paper where we were flabbergasted by the finding that even for the discrimination between low grade and high grade, not every single pathologist agrees. So if you don't agree with the most basic thing, how do you expect us to treat adequately patients? And so there's unmet need, but still, I don't think it's up there. With respect to the surgeons, many, many studies in the past, EORTC here, the RESECT study have clearly shown that there are differences between surgeons. But again, whether a robot will replace Ashish Kamat operating tomorrow, I'm not sure, certainly not today. And so what are we left with is the risk prognostication. And I had shown at that time a study led by Jethro Kwong in Toronto and my partner Girish Kulkarni, where we assembled a huge cohort, probably the largest, to the best of our knowledge, close to 13,000 men from North America. You Ashish and the MD Anderson were part of the study as well as many centers in Europe, trying really to improve on the risk prognostication.

And we know that we need better tools. We know that the EORTC risk calculator was completely outdated and was using the old 1973 classification. CUETO brought in the BCG treated patients. But same thing, the 1973 and the EAU, although it does include patients who had the new low versus high grade classification, only consider patients who are not treated by BCG. So there was a huge room for improvement. And to do so together with all these different institutions, close to 30 North American and European institutions, we assembled a training and an external validation set where we compared AI to the gold standard, that would be the EAU risk calculator with the outcome, which is the progression to muscle invasion or metastatic disease. And we used many parameters to analyze the data. And the bottom line here is that you can see that whatever way you looked at the data, the AI clearly outperformed the other models and the EAU risk calculator.

And not only in terms of C-index, but also in terms of calibration and decision analysis. And then ultimately, with respect to some of the groups which truly need a very fine-tuning. And the group of the IBCG, as you know, has done a lot of work to improve the classification, the intermediate risk. AI was actually truly able to fine tune this risk. On the left-hand side, you see that using the IBCG factors, you can actually discriminate between two groups. But using AI on the large data set, you could see that a lower group has exactly the same outcome as low risk, that the highest group has a risk of progression, very close to high risk patients, and there's a real true intermediate risk which is different from the low and the high risk group.

And so I think the summary is that this paper, which is under revision and hopefully will be accepted soon, and where the AI model will be available to be used by any patient, any physician worldwide is ready to go. It did outperform the existing risk calculator. It will probably inform many guidelines about optimizing the risk management. And interestingly enough, it included patients who were treated according to guidelines, patients who never received treatment and a long-term follow-up in more than 3000 patients where it also actually capture long-term data. So I really think that AI and risk calculation is here to stay.

Ashish Kamat: Thanks so much, Alex. And again, AI clearly can take data that we have and make a lot more sense out of it and come up with its machine learning algorithms and refine things to get patterns that we don't actually see. But ultimately, a lot of the data that comes into the AI models comes from what we put in and one of the key parameters is what our pathologists put in. Right. So love to hear from Hikmat on his perspective. Hikmat?

Hikmat Al-Ahmadie: Yeah, thank you for this opportunity to present this aspect from the discussion. And thanks, Alex, for providing the clinical background or the clinical parameters for the risk stratification. What I will not tell you in 2025 is that everything can be solved by histopathology alone. What I would also tell you is that there's a benefit from not throwing away everything that is based on histopathologic evaluation and perhaps think of it as how can we improve it with the current technology that is provided by image analysis and AI. And that is for an important reason. We rely on histopathologic features or histopathologic evaluation of tissue or in bladder in particular to provide diagnosis, provide staging of the disease, provide important clinical parameters that link nicely or correlate nicely with important clinical parameters such as recurrence and progression.

And this has been done in many studies. This is one example. Well-done histopathologic analysis can provide important risk stratification and classification. To the point that we use this as was alluded to, we use the histopathologic information in pretty much every risk stratification and every guideline, national and international to incorporate and provide risk stratification that can help us manage patients with bladder cancer.

Of course, there is an inherent problem with the classification because they're histology based, they're subject to subjectivity, but that could be related to more than one factor. One, the biology of the disease is not that well-categorized. The disease happens at a spectrum of morphologic features that makes it difficult to provide categories most of the times and or histologic features can overlap, which add to the difficulty on establishing categories that can, in which you can classify every patients. There's a lot of patients that fall nicely into these categories, but there's going to be always overlap. And as a result, we know the difficulty and that it comes with this by the fact that we have many classification systems in urothelial carcinomas that don't necessarily align with each other. But there's a lot of overlap as well.

But beyond just simple classification, there are many other histologic features that we can rely on to help us in risk stratifying and managing bladder cancer patients, including sub-staging or quantification of lamina propria invasion, the depth of overall invasion, histologic subtypes, which are relevant in all stages of urothelial carcinoma and lymphovascular invasion. These are all important histologic features. And as you could imagine being distinct histologic or microscopic features that can be seen on a slide, they lend themselves naturally for any image analysis application. And this is why this can also be an important feature that can be used for any of the AI models that can be developed. And there are things happening, which we're kind of behind in the wide applications of image analysis in urothelial carcinoma. But there is some evidence to show that this works. And investigators have used two approaches.

One is try to use image analysis or AI in recapitulating and improving refining current classification system, like the example, the article on top where you can provide more accurate, more reproducible grading of urothelial carcinoma. Or you can have a different approach that is not necessarily relying on existing histologic features, but extracting features from the slide and try to correlate them with outcome or responses to different treatment strategies. Both have pros and cons and both can provide relevant information. And we need more of these studies to help us improve what we do for bladder cancer.

So overall, we know why AI can improve histopathology classification and risk stratification in urothelial carcinoma because we know histology classification has been there for a long time. It's practical, it's cheap, it's in every hospital and every laboratory. It provides information that is really reflective of underlying biology and it's flexible and it has been shown to be like the backbone of many of the newer molecular classification systems by layering this additional information on the histologic basis. And with the current wide use of slide digitization, which is available now in many, many small and large healthcare facilities, there is really a great opportunity to capitalize on this and use this resource to try to improve the classification and risk stratification of urothelial carcinoma. So as a take home message, yes, histopathology classification is not outdated completely, but it could use some improvement by utilizing the current or future image analysis or AI tools to help improve that process. Thank you.

Ashish Kamat: Thanks so much, Hikmat. Again, you highlighted the field of AI and we say AI, but it's actually machine learning, right, when it comes to essentially pathology and the field of pathologic interpretation. And then Alex, you showed so well how AI and machine learning can take the clinical parameters and pathologic parameters and put them all together and help us manage our patients. One of the things that we all know, everyone that's been doing this for a while is that bladder cancer is an extremely complicated complex field and it is constantly evolving, right?

So I'm going to ask you Hikmat first, again, if you have a crystal ball and you're looking into it, how do you think that these tools, and I like the way that both of you framed it, these are tools to help us. These are not going to replace us. I don't think anyone needs to get on the bandwagon that, oh, we need to be afraid of these. But how do you think these tools are actually going to be implemented in your workflow, Hikmat, on the pathology side, or when the surgeon's giving you a specimen, how do you think this AI is going to be implemented in say, frozen sections or telling us in real time, do I need to get a deeper resection on a TURBT, et cetera, et cetera? A little insight there.

Hikmat Al-Ahmadie: Yeah, I mean, I think it's not there yet, but there are ways, just by the way, knowing how it works. And there are examples, for example, in the breast and in the prostate cancer, there are AI algorithms that can detect the presence of cancer. It can detect the presence of lymph node metastasis. I don't think there's anything yet there in the bladder, but there's no reason why this cannot be developed for the bladder. If you can provide enough information to tell you carcinoma in situ, for example, versus benign, and you can have a quick section that can be fed into, made into a digital slide and fed into an AI algorithm at the frozen section margin, there's no reason why it cannot detect that abnormal histologic finding provided that you trained it on the frozen section slide, not on a permanent slide.

So that is the key. They have to be trained well to detect what has to be detected, and that distinction between benign, malignant. And then there's no reason why it cannot perform well. The same thing would be for a deeper TURBT. If it doesn't detect muscularis propria, if it's trained to detect muscularis propria or identify it, then it will flag that case, that yeah, this case doesn't have muscularis propria.

Ashish Kamat: And Alex, again, you showed how it helps you sub-stratify patients and risk identify patients, but that's using AI in a broad range of risk stratification. Right. Now, Alex Zlotta has been doing this for years. You have insight into bladder cancer that I'm sure even you don't understand how you have that sixth sense. And we've seen in previous iterations with IBM Watson and others that we and Memorial have worked with, that the tools just don't get there when it comes to the top, say 5% of clinicians. So I would put you in the top 0.5% of clinicians, Alex. So how would you say AI would help you? Not the average person. I'm sure it can help the average person, but how will it help you, Alex Zlotta, the premier bladder cancer expert?

Alexandre Zlotta: I'm going to tell you honestly Ashish, that our profession is one of the most humbling, and I don't know about you, but I actually feel sometimes that the more I know, the less I know after many, many decades. And back to the same question. The second thing is the number of times when if you review your own surgeries and this and that, and then you look at what you did and say, oh my God, how stupid could I have been? So I think for all of us, we all think that we're the best surgeons on earth. And then when you look at the data, we didn't do this, we didn't do that, and we didn't do that. So there's a huge disconnect. We don't even use guidelines. Look at the number of people who are not even treated according to guidelines when they should.

The number of people who don't get repeat TURBT. So what I'm getting at here is that we're going to go baby steps by baby steps. I think that AI is one of the steps to highlight and to flag patients where instead of just having the gestalt, Mr. Smith comes into your office, oh, I think that your progression rate will be around 10%. And then you put this risk calculator based on so many parameters, and you realize that the risk of this progression is about 22%. Everyone most likely will have a completely different discussion with Mr. Smith and his wife and the family if the risk of progression is over 25% or 20% and less than 10%.

And so that's how I see those baby steps on the daily basis. To start with, I will simply say that progression is only one piece of the puzzle. There's going to be recurrence. And I think the beauty about all these trials is that we can start feeding these models. Like it's not that Google stops working. They feed data every day and every day and every day. And then if we keep feeding the models, we'll refine those models for the next generations.

Ashish Kamat: Excellent, excellent point. And I think that's the critical part. Right. We don't want to look at it like we have done in the past with some of these large multi-million dollar efforts to see if we can actually replace the clinician. This is to help the clinician take care of the patient, both when it comes to pathology or on the clinical side, or even in cystoscopies, right? There are AI models being developed where in real time it can look at the image on the screen and tell the urologist, hey, look at that more carefully. That looks like cancer, even though you may not have biopsied it or something like that. So again, in the interest of time, we'll stop now, but this is an evolving field. I'm sure we'll have both of you back again on the forum in six months or a year to talk about the next iteration. But for now, I want to thank you both for your time and spending it with us here on UroToday.

Alexandre Zlotta: Thanks.

Hikmat Al-Ahmadie: Thank you.