Zachary Klaassen: Hi, my name is Zach Klaassen. We are at SCU 2025 in Cartagena, Colombia. I'm delighted to be joined as always on UroToday with Dr. Matt Cooperberg, urologic oncologist at UCSF in San Francisco. Matt, thanks for joining us on UroToday.

Matthew Cooperberg: It's a pleasure to be here.

Zachary Klaassen: So you gave a talk at SCU looking at artificial intelligence in prostate cancer, which probably could be its own symposium on itself, but maybe just talk about how AI has taken over our life and really how it's moved into medicine over the last several years.

Matthew Cooperberg: Yeah. I mean, AI is everywhere.

Zachary Klaassen: Everywhere.

Matthew Cooperberg: It's a 15-minute talk and we're supposed to cover screening, diagnosis and treatment. I've focused mostly on diagnosis, but it is coming everywhere. I mean, maybe not screening yet, but diagnosis, absolutely. Imaging, pathology, those are ground zero today. And treatment is only a short distance away. And that's just prostate cancer. Right?

Zachary Klaassen: Yeah. Absolutely.

Matthew Cooperberg: I actually started with the slide looking at the lay of the land in terms of FDA approved AI tools in anything up through therapeutics across medicine, and you can picture what the shape of that curve looks like. Two tools approved in 2017 and this year it was 28. As in everything else, prostate cancer is lagging way behind breast cancer. These data were already completely out of date. When the paper was published two months ago, there were seven or eight approved for prostate cancer and I think already 30 something for breast. And there's already more than when the paper was published in prostate.

So things are moving incredibly, incredibly quickly, and I don't think anybody can even predict really how fast things are going to move, and in exactly what directions. Some of the things that we're seeing now were predictable. I think some are not and I think what's going to happen over the next five, 10 years-

Zachary Klaassen: Crazy.

Matthew Cooperberg: I know. It's remarkable.

Zachary Klaassen: Tell us a little bit about some of the highlights of that screening and diagnosis, because like you said, there's a whole bunch of literature out there. We've got biomarkers, et cetera, MRI. What were you focused on for your talk?

Matthew Cooperberg: I mean, first of all, you're absolutely right that we already have a saturated space. The number of markers that we have for every decision point from do you check a PSA? Do you do a biopsy? Do you treat the cancer? How do you follow up? How do you do active surveillance? How do you manage advanced disease? We've got more markers than we know to deal with at each of those decision points, including imaging tests, which are in that marker space.

So when we talk about a pathology AI tool, for example, well, it's playing against Decipher and Oncotype, et cetera. And of course in the pre diagnostic space, the conversation is really about MRI. And to an extent PSMA PET and how AI is modifying our use of those tests. So I spent a lot of time on MRI and on pathology because that's where most of the action has been so far, and the action is all the last couple of years and it's moving incredibly quickly.

I think one of the coolest things is that in both the case of MRI and in the case of pathology, we've seen these big open source efforts, the PI-CAI initiative for radiology and the PANDA initiative for pathology, which are really making a lot of news. And to be clear, there's of course a lot of companies in the space. We can talk about them too.

Zachary Klaassen: Sure.

Matthew Cooperberg: But this grand challenge framework that they've used for PI-RADS and for radiology and for path, I think is really, really cool because it shows the power of what we can do in a digital era where you don't need to send physical slides around, you don't need a lot of genomic expertise. They literally just put thousands of biparametric MRIs on the internet, pathology images on the internet with the outcome and told the world to go at it. And in each case, I think for PI-CAI, they got hundreds. For PANDA, they got over a thousand submissions from literally dozens of countries, which is also really cool.
These are global efforts and they then took the top five, which in the case of PI-CAI just so happened to be five different countries, which I think is also really cool.

Zachary Klaassen: That's awesome.

Matthew Cooperberg: And there's this math, I don't understand where you can merge the five. It's this Netflix style thing where you can take five mathematically distinct models and merge them and come up with your best model, aggregating the best of each of the best and that best model in the case of... In the case of pathology, the goal was just to predict accurately a consensus call on the grade from the pathologist, which was able to do with Kappas in the '80s.

Zachary Klaassen: Wow.

Matthew Cooperberg: In the case of radiology, it was a bigger deal because that study was powered for non-inferiority, but it actually won.

Zachary Klaassen: Wow.

Matthew Cooperberg: So they had several dozen well-trained GU-focused radiologists reading out these cases. And the goal was not giving a PI-RADS score for that one. The goal was predict grade 2 or higher cancer, and the AI beat 75% of the radiologists. This is version one and it's free. This is open source.

Zachary Klaassen: It's only going to get better with more data.

Matthew Cooperberg: Well, exactly. This is version one. So there was just a follow-up paper on the PI-CAI 1 where they looked to see, "Well, can this improve a human radiology workflow?" So the answer was yes. They looked at the accuracy for the radiologist doing reads. Then the radiologist plus PI-CAI did a better job.

Zachary Klaassen: Wow.

Matthew Cooperberg: But they still didn't do as good of a job as PI-CAI by itself.

Zachary Klaassen: Wow.

Matthew Cooperberg: Obviously from a regulatory standpoint, we're not replacing radiologists-

Zachary Klaassen: Sure.

Matthew Cooperberg: ... anytime soon. But if you think about the community radiologist that reads a bunch of mammograms in the morning and maybe picks up a chest X-ray and then takes a multiparametric MRI off the queue after lunch.

Zachary Klaassen: At 3:00 PM.

Matthew Cooperberg: Right. I mean, there's no way that person three years from now should be doing that work on assistive. Maybe even next year. Maybe three years is too conservative because this is all going to move so quickly. The regulatory piece will take a while. And literally last month there was a paper in radiology trying to put some parameters around how we should be evaluating these tools. So checklist score for AI tools in radiology, which is really important.

Zachary Klaassen: Sure.

Matthew Cooperberg: We need this sort of frameworks.

Zachary Klaassen: I want to spin off on something you said, the global effort. I think when we look at... Certainly there's parts of the United States, and I'm sure parts of Colombia where there's underserved areas, but we think about really third-world countries that could eventually access to digital pathology, send it to an AI platform. I mean, this could really revolutionize how we're diagnosing and eventually treating prostate cancer.

Matthew Cooperberg: Absolutely. Actually the pathologist in the news doing a transoceanic surgery, and that's obviously really involved. Sending the pathology image across the ocean-

Zachary Klaassen: Super easy.

Matthew Cooperberg: ... that's nothing, right?

Zachary Klaassen: Yeah.

Matthew Cooperberg: And you don't need to send tissues around. You don't need to do RNA extraction. As much as the genomic story has been fantastic over the last 15 years, access to Decipher or Oncotype or Prolaris outside of the US and selected parts of Europe is still really, really limited.

Zachary Klaassen: Sure.

Matthew Cooperberg: Sending a slide and doing an extraction, it's not the end of the world, but you have to mail it. We have to deal with customs and all that. Sending a digital path image is basically instantaneous and therefore much, much lower cost. They don't need to buy an [inaudible 00:06:47] or even RT-PCR reagents. I think we have a while to go until we say, "Well, the accuracy is equivalent and we know as much about these as we do about the genomic tests." But on the other hand, they can evolve much faster.

Zachary Klaassen: That's right. That's right.

Matthew Cooperberg: It's really interesting.

Zachary Klaassen: Tell us a little bit about the treatment, because I know you said earlier too, this is early days for AI and treatment. Maybe I allude to some of the things you touched on that part of your talk.

Matthew Cooperberg: Yeah, this is much sooner.

Zachary Klaassen: Sure.

Matthew Cooperberg: Much, much further upstream, but it speaks to the fact that we need to have our eyes on this ball. And I said in my conclusion, in all seriousness, I would not recommend my children going into pathology or radiology today.

Zachary Klaassen: Sure.

Matthew Cooperberg: Unless they're really into AI development and they have a plan to be at the leading edge of the way this is all going to evolve. And frankly, primary care is probably not so different. I think in therapeutic niches, while this is going to be AI-assisted, so we're already seeing some studies on AI helping with complex oncology algorithms. There was a paper very recently for prostate automated contouring of the prostate for radiation oncology. And this is obviously coming as radiology goes. I used to think I was fine in surgery for the rest of my life. I think I'm fine in surgery for the rest of my career because I've gotten enough gray in my beard.

I think for the residents coming through now, I wouldn't be so sure. I mean, first it's going to be augmented reality, and this is coming. We've already done some studies at UCSF and elsewhere that Peter Carroll has been leading using PSMA-fluoro, diagnosis of lymph nodes. We have an ARPA-H grant to develop a much higher resolution intraoperative diagnostics of where the nerve is, et cetera.

Zachary Klaassen: Cool.

Matthew Cooperberg: But this is still just sort of improving. What we can see a layer on top of that is you'll let the computer help guide you and say, "No, you really shouldn't cut that." New da Vinci 5 has this weird haptic thing that we all kind of hate, but maybe eventually we'll actually not let you cut the ureter, which I think those of us that have been driving for a while and hate the cars that try to keep you in the lane, at least it's going to be like that for people who've been doing it for a while, but a new generation coming in, this will just be part of how you learn.

Zachary Klaassen: That's right.

Matthew Cooperberg: But then eventually Hopkins just had a model that could take out a gallbladder in a box. And the scary thing about that one is they did not program it to take out the gallbladder. They just showed it a bunch of videos of cholecystectomies and then set up this ex-vivo model. This was just a box. But just by watching videos, this da Vinci looking robot could identify, dissect out and take out the gallbladder. So it is a matter of time.

Zachary Klaassen: It's a matter of time.

Matthew Cooperberg: I don't think that is coming clinically in the next 10 years, but 20 or 30, who knows?

Zachary Klaassen: Who knows?

Matthew Cooperberg: It's all changing. I saw an interview literally just this morning from some Google higher up who basically said nobody should go into medicine anymore.

Zachary Klaassen: Wow. Really?

Matthew Cooperberg: And obviously that's an extreme statement. And the problem that that statement highlights is the same thing happening everywhere else like in tech and programming. At least so far, you still need people with experience and wisdom guiding what these things are going to do. And I think the real concern is, "Well, how are you going to cultivate the next generation?" So in tech, all the senior programmers love GPT because it will churn out pages of code for them, it does a great job, and they have this little Copilot. And you don't need to hire six junior programmers anymore.

But if you've never been a junior programmer, you're never going to become a senior programmer. And how do you get good? And if 80% of the surgery gets done by the robot, how do you build up the instincts to realize when something is off program that the robot might miss to take over? It's the same thing with the driving, right? If you're in Waymo 80% of the time, you don't get the instincts to react to something that is not typical and avoid something that the robot might miss. And we're heading into this era of dependence on these things that nobody really knows what's going to happen.

Zachary Klaassen: It's the wild west. We used to think-

Matthew Cooperberg: It's the wild west.

Zachary Klaassen: We used to think PSMA PET, that was the wild west. AI is the wild west.

Matthew Cooperberg: It is. Oh, it really is. It really is. This is wild west. This is like they just found gold.

Zachary Klaassen: That's right.

Matthew Cooperberg: And people are starting to load up the wagons. We have no idea what they're going to find when they get there. It's something to really keep our eyes on. And I do think leaders in medicine need to be driving these conversations. It really cannot be the tech companies driving these conversations. Obviously they are very excited. San Francisco is in a brand new tech boom. Exactly around these AI startups. It's all these 20-year-olds co-living in these giant houses in this valley changing the world. And I think the question is, do we have anything to say about the ways in which the world changes? Some of that are going to be incredible. Some of that are going to be concerning. And I think we need to be at the head of that conversation.

Zachary Klaassen: Sure. Last question. From a very realistic standpoint, let's say next three to five years, how do you see AI pass regulation, everything else in the clinic for prostate cancer?

Matthew Cooperberg: Well, the tools that are coming out. The industry tools are coming out. So Artera is already there, right?

Zachary Klaassen: Yup.

Matthew Cooperberg: I mean, I've never seen a test go from presentations to NCCN approval-

Zachary Klaassen: To zero to 60.

Matthew Cooperberg: ... so fast. I mean, it's incredible. And that speaks to their really unique access to these RTOG trials, some really good academic leadership that they've had. And Felix Feng's contributions in particular. So that's an incredibly fast-moving story. And they're already looking at other cancers and all this sort of thing. So Artera is clearly example one. There's already two other companies out there coming fast. I think Miromatrix and Pathomic. There's others in the pipeline for pathology. There's lots of companies in radiology that are going to try to get into this space.

How Artera evolves, getting to this question about the genomics that's exactly I think the question we're going to have going forward. So they're right up against Decipher. All right, but is that a competition? Is that actually maybe a complementary question? I'm sure there are things, and I actually have an R01 to do exactly this sort of question. What can we learn from the genomics in parallel with the AI? Can genomics improve the explainability of the AI tools? What is it that we learn from AI that we're not learning from genomics and vice versa?

All the genomics companies are partnering with pathology AI companies to some extent or another. So there's going to be a really interesting convergence there that I think is going to improve our fundamental understanding of the biology that much more. We can go back to the what is cancer conversation too.

Zachary Klaassen: Sure.

Matthew Cooperberg: I think actually that collaboration between genomics and pathology will really help us understand actually what cancer is because we are getting to a... I don't want to say a post-pathology era, but an era where the definition has to be more than just H&E pathology. So that is all really exciting and parallel things are happening in radiology. There are tools that are hitting the clinic as we speak, or we'll be hitting the reading room, I suppose, not the clinic-

Zachary Klaassen: Sure.

Matthew Cooperberg: ... as we speak, that will help radiologists or eventually you could imagine where this is just in the urologist's office and contours your tumor before the biopsy on the spot.

Zachary Klaassen: So it's right on what they're planning.

Matthew Cooperberg: Exactly. I mean, there's these companies that are trying to do this sort of micro MRI in office sort of thing. So we're going to see a huge competition for technologies. And again, it's our job, and this will be for the HSR set to really figure out where is the value in all this because equity of access has been a huge problem throughout, and what we don't need is more proton centers.

Zachary Klaassen: That's right.

Matthew Cooperberg: Or more robots. We don't need everybody buying the technology. We need to figure out ways to get these in front of patients who have had no access to any tools so far. I think it will be easier, as you said, with the AI tools than has been with genomics and some of the prior technologies.

Zachary Klaassen: We covered a lot of territory, Matt.

Matthew Cooperberg: Oh, good.

Zachary Klaassen: Always appreciate your generosity with your time and great discussion on AI. It's going to be an exciting next couple of years for sure.

Matthew Cooperberg: It really is. Watch the space.

Zachary Klaassen: Watch the space. Thanks, Matt.

Matthew Cooperberg: Great, thanks.