Read the Full Video Transcript

Zachary Klaassen: Hi. My name is Zach Klaassen. I'm a urologic oncologist at the Georgia Cancer Center in Augusta, Georgia. I'm delighted to be joined on UroToday by Dr. Nick Kavoussi, who is a urologist at Vanderbilt University in Nashville, Tennessee. Nick, thanks so much for joining us today.

Nicholas Kavoussi: I'm very happy to be here.

Zachary Klaassen: So today, we're going to discuss your awesome SUO 2024 presentation, artificial intelligence in RCC diagnosis and treatment. I'd love for you to walk us through some of the highlights of your presentation.

Nicholas Kavoussi: Absolutely. Yeah, happy to talk about artificial intelligence in kidney cancer diagnosis and treatment. And I think to really start, we need to talk about artificial intelligence. It's a hot topic these days. And really, what it describes is any technique that allows machines to mimic human intelligence. What this really means is that AI is about prediction. The same way we can make predictions about the world we experience, we're teaching computers how to do that. And AI models are typically data driven compared to conventional statistics, which are model driven. And this allows the models to be flexible for fitting in the data that they're representing.

It comes at a cost in terms of interpretability. We don't always know what these models are doing, which can make them challenging to understand, improve, and figure out how to integrate these models into our clinical practice. But it really has opened up our ability to take a variety of different data types and study them and understand how these data types predict outcomes in ways we haven't before. And this could be anything from imaging, pathological data, surgical evaluation of the surgical scene, as well as kinematic data from a robotic chain or elements of the EHR data.

To develop these models, what you do is you take these data types, you associate these data types with an outcome, and you tell the computer, "OK, in these instances, this is what we found. And in these instances, we found this other variable or this other outcome," and do this in a recursive way such that you can actually predict and make smart tools that can help us make better decisions for our patients in the operating room.

And there's a need in kidney cancer, and there are a lot of different reasons why that's needed. But really, we need to improve the way we select patients for treatment. We need to better risk-stratify patients and understand how patients are going to respond to our treatments to get durable and safe outcomes for our patients. And I want to focus mainly on renal mass classification for two reasons. One is that I think this is where the most robust studies have been done. And two, I think it really represents what's going on in artificial intelligence in the clinical sphere overall. And I think we can get a lot of lessons from where we are based on what people have been doing with renal mass classification.

This alludes to studies where people are taking preoperative imaging and associating the imaging with postoperative pathologic grade, stage, outcomes to try to understand what we can see on imaging that might predict how patients will do after surgery, after treatment, whatnot. There are a lot of studies that look at renal mass characterization, and they do fairly well overall at first glance. And these studies have looked at subtype, staging, grading, and, as I mentioned, prediction of recurrence.

However, if you really look at these studies, there have been big issues in two realms for the artificial intelligence-based tools in kidney cancer. That is based on external validation or comparison with experts. And these are critical steps to understand where these tools can be used to help us take care of our patients and improve our practices. I'll give you an example of the largest study done to look at pathologic prediction compared to experts of renal tumors.

This was done with over 1,000 MRIs across five institutions from the group at the University of Pennsylvania. They compared a bunch of different AI-based tools—models from different sequences of the MRI—and looked at those in comparison to experts and other deep learning models. And what they found is that their tools, overall, are comparable to experts. So even in the most robust, diligent case of developing these models, we're not making anything that's better than experts at this time.

Now, hopefully in the future, we will be able to, but there are potential applications. Still, we can potentially use this to improve access to interpretation of MRI models in places that don't have access to a dedicated radiologist. We might be able to improve reproducibility of models and improve efficiency of reading these MRIs. So there's a lot of potential there but still limited by the black-box nature—not knowing what these models are doing all the time—which prevents us from really understanding where we are with them and how to make them better.

There's a lack of data standardization in terms of what we need to use and train these models efficiently, and we haven't really validated these studies either externally or against experts, which I think is necessary to understand where these are going to be applied. And I think because of this, we're still pretty early in the hype cycle in terms of application. But I believe these tools are still going to come as we build better and better data sets and models and allow us to evaluate things we couldn't do previously. And it's our responsibility to evaluate and usher these tools in responsibly to help improve our practices and take care of our patients. So again, thank you for having me and happy to discuss further.

Zachary Klaassen: Nick, thanks so much for highlighting your SUO presentation. There's a lot we can unpack in this discussion. In your presentation, you talked about diagnosis. I know at SUO, we talked about treatment and artificial intelligence. There's a lot of options here. It's early days, as you mentioned. I'm going to put you on the spot. If you had one clinical unmet need today that you could have AI validated in clinical practice, what do you think that would be at the moment?

Nicholas Kavoussi: Yeah, I mean, I think the biggest unmet need is having those patients come in with these incidental renal masses and be able to say, "Hey, you're fine, get out of here. Don't worry about it," or, "Hey, this is something that's going to cause you a problem." If we could do that, that would be amazing. But we're just not there.

Zachary Klaassen: Yeah. No, I think too, as we look over the next three to five years, we've certainly seen AI in prostate cancer with the Artera AI prostate test. We'll probably get there with kidney cancer. One thing I've thought about is if we had a predictor for who would need a year of adjuvant pembrolizumab after a high-grade operation for nephrectomy—KEYNOTE-564 data—there are a lot of things that we can look at over the next three to five years. What landscape in the next, let's say five to ten years, could you see this really being in the workflow from beginning to end of management?

Nicholas Kavoussi: Yeah. Well, I think there's a few things to consider. I mean, I think we still have pretty unmet needs in developing these data sets. I think we need large and accessible data sets to accurately and appropriately train models for specific tasks.

Zachary Klaassen: Yep.

Nicholas Kavoussi: And I think that there are two real needs for these data sets, no matter what you're studying—these incidental masses, what you're going to do, adjuvant therapy, what have you. We need these data sets to reflect the heterogeneous nature of renal masses and kidney cancers to allow for robust clinical application of these AI models. And, two, the data sets need to be curated toward accessibility and consistency, because we live in a world with many different EHRs, many different imaging programs, and we want to be able to consolidate this information so it's useful.

Specifically to your question, I think the application of AI in the clinical sphere will be gradual. I think over the next five, ten years, we're going to see really small tools to implement these models into clinical practice. I think as a society, we're more comfortable with AI; our patients are more comfortable with the idea of AI. So we're going to see things like AI-based risk calculators and gradually move towards guidance in diagnosis and surgery. And I think at this point, it's important to mention the FDA has a white paper regarding regulation and evaluation of these technologies. So there's going to be specific standards that these tools are going to have to be held to, which is important for safety and application.

Zachary Klaassen: No, great answer. I think too, as we look at 2025 and beyond, we should have girentuximab PET imaging coming down the pipeline, hopefully soon too. So that'll be a whole other data set and avenue for discussing with patients. It's early days in RCC, but there's a lot of unmet needs. It's exciting to see your work and the work you guys have put into it. So we'll be keeping close tabs on that. Any take-home messages for UroToday listeners today?

Nicholas Kavoussi: Yeah. I think overall that we need to acknowledge AI is here in some sense, and it's coming in other senses. And we've got to familiarize ourselves with how it works and what it can do.

Zachary Klaassen: Yeah.

Nicholas Kavoussi: The current studies are limited. They're limited in what we can do for our clinical practice. But we're going to have better studies come out. We're going to have better tools come out. It's our responsibility as clinicians to be the gatekeepers for this technology and make sure it's safely and appropriately used for our patients.

Zachary Klaassen: Awesome, Nick. Thanks so much for your time and expertise. Congratulations on a great presentation and for your time on UroToday.

Nicholas Kavoussi: Hey, thank you for having me and enjoyed being here.