Nathan Perlis: Absolute pleasure to be here. Thanks for including me, Zach. We love following everything that you guys talk about.
Zachary Klaassen: Awesome. Well, this is really something near and dear to my heart because I think we all deal with this in the clinic. And I think before we get into this study, maybe just highlight, at a high level, the PI-RADS scoring system, specifically why PI-RADS 3 can be a bit of a clinical dilemma.
Nathan Perlis: The PI-RADS scoring system is a system that has been designed and validated across the world, and it basically is used by radiologists to score spots that they see on patients' prostate MRIs. And the idea is that the higher the PI-RADS score, the higher the likelihood that there will be clinically significant prostate cancer in that spot if the patient should go for a biopsy of that lesion. The scoring system relies on the way the image looks on various sequences on MRI. And the reason why the 3 is the trickiest area is because PI-RADS 1 and 2 are felt to basically have a low chance of cancer, and they're usually not biopsied. 4 and 5 are more distinctly worrisome on imaging and have a much higher likelihood. But PI-RADS 3, they call it intermediate. And there have been many studies looking at clinically significant cancer rates if you target a PI-RADS 3 lesion and several meta-analyses. And the clinically significant cancer rate in a PI-RADS 3 lesion is so all over the board. It can be as low as two to 5%, or even as high as 50%. The challenge being is that so many different inputs lead up to a radiologist calling it a PI-RADS 3 and eventually finding it. So one radiologist would call it one thing, another radiologist would call it another.
Sometimes in your specific patient population, you might have a lot of clinically significant prostate cancer, or less. And so there's so many different subjective inputs to PI-RADS 3, that if you just rely on the bigger publications that comment on what should be done in PI-RADS 3, you may not see the same output in your own patient population. So a lot of clinicians have a difficult time knowing what they should do with that PI-RADS 3 lesion found on a prostate MRI.
Zachary Klaassen: Yeah, that's a great background. I think you nailed it. I mean, I think it's 50%, 40%, 60%. It depends what percent you're in for that PI-RADS 3, right? So it becomes a problem. You guys developed a local clinical risk model and you cross-validated. That's the topic of this study. Maybe just set up the study design, the methods that you guys used.
Nathan Perlis: Absolutely. Thanks for that question. So basically, we have several large academic sites here in Toronto, and we took the two biggest ones with large MRI databases. There were roughly 3,000 patients from each of the sites. And what we did is we narrowed all those patients down to those that had no previous diagnosis of prostate cancer, recent PSA, so that they had to have full clinical data, and that they had a targeted and systematic biopsy. And we took the first group of patients and developed a clinical prediction model, internally calibrated it, and then what we did is we used that model on the validation cohort, which was from the other hospital, to see how well it would predict whether the patient would have clinically significant biopsy on their PI-RADS 3 lesion fusion biopsy. The fun thing we did is we compared that to everything else that's out in the literature, from the most complex models like the ERSPC and the Radtke model from Europe, to the most basic way that a clinician might use to consider whether they should biopsy a PI-RADS 3 lesion, like using a simple PSA density cutoff or an ADC map cutoff. And that's generally what we did in this study.
Zachary Klaassen: That's great. And maybe just walk us through the high-level results. I know you guys, as you mentioned, compared it to these other models. How did it perform?
Nathan Perlis: So it performed really, really well. Our model, again, it was calibrated in Toronto and then tested in Toronto, performed the best. So it did better than the other multimodal models, and it did better than the simple cutoffs. And at the end of the day, if you were to rely on our model for your decision-making, you could avoid roughly half the biopsies on PI-RADS 3 lesions that you otherwise ... If you would've had a decision model being biopsy everyone with a PI-RADS 3, if you would've used ours, you could have avoided half of those. So it was really quite impressive. Interestingly, the simple cutoffs, like the PSA density of less than 0.08, outperformed some of the other models out of Europe. So that was another really interesting finding in our study.
Zachary Klaassen: No, that's great. When we think about these ... And in my clinic, PI-RADS 5, they're all getting biopsied. The majority of PI-RADS 4, unless they're elderly, comorbid, on blood thinners, are going to get biopsies. It's the 3s that are the problem. And so with your model, you're safely avoiding half the biopsies. How do we take this information from your guys' study to the clinic tomorrow when we're counseling patients with PI-RADS 3 lesions?
Nathan Perlis: I'd love to be able to say, "Just plug in our model and you're going to get just as good outcomes as we do," but unfortunately that doesn't reflect our findings. I think that ours did so well in our local environment because it can take into account referral patterns. It can take into account expertise from the GU pathologists, the GU radiologists, the people doing the fusion biopsies, and it can also take into account just the prevalence of clinically significant prostate cancer in our locale. Here in Toronto, we have a multi-ethnic population. We have our own unique settings. That being said, the items in our model are available everywhere. So we used PSA density, we used age, we used prior negative biopsy. These are all entities that you have available to you. So if you are able to internally validate and calibrate our model or even the ERSPC model to your own setting, that would be the most ideal way to do it. And barring that, we do show that using a PSA density cutoff, which is something that's available to everyone very easily, is probably the next best thing. So I would say that in the 90% of centers using a PSA density cutoff, for us, it was 0.08. Some other studies have used 0.1. It seems basic, but actually it pans out very well across many studies, and that's a good place to go. Other than that, I think we need to find better ways to calibrate these models to our local institutions and our local populations.
Zachary Klaassen: That's great. Very actionable. I want to ask about future work with your guys' model, but I'm going to put you on the spot a little bit. There's been a lot of data coming out, IsoPSA MRIs. Is there thoughts about incorporating that? Maybe lead off into the future work aspect.
Nathan Perlis: So I think that there is a lot of room for incorporating, first of all, specific entities from MRI. AI can very easily generate much more specific outputs around ADC, around diffusion-weighted imaging, around prostate size, various contouring. So we'd like to, and we started to, think about incorporating that into our models. Of course, genomic classifiers, IsoPSA, these are all things that can be added. It's a fine balance between trying to find something that's affordable, clinically meaningful, and simple, than throwing the kitchen table at every patient. So we do struggle with that. And we were pleased to see that calibration works, but not everyone can calibrate. And aside from that, outputs from MRI work and outputs from PSA can work too, and not everyone with a PI-RADS 3 lesion needs to get a biopsy.
Zachary Klaassen: I think you nailed it. Simple is going to flow in the clinic. It's going to be incorporated into the workflow much easier, for sure. Nate, fantastic discussion. I'm glad to have you on to discuss this. I know a lot of us listening to this will be dealing with PI-RADS 3s probably in their clinic this afternoon. Anything we haven't hit on? Any take-home messages for our listeners?
Nathan Perlis: I can't talk about MRIs without encouraging everyone, clinicians, residents, fellows, look at the images yourself. You can plug it in and you can say, "Oh, but my radiologist isn't good at this," or, "The GU pathologist called this." You're the clinician. You're the quarterback. Know what a PI-RADS 3 looks like. Not all PI-RADS 3s are created equally, and that comes up every day in clinic with me. So look at them. I think you'll get enough gestalt from the image to decide what needs to happen, and the patient and the PSA density. The struggle is real, but there is a lot of data coming out to show that if we rely on our clinical parameters, we can help patients because not everyone needs an MRI fusion biopsy on a PI-RADS 3, but not everyone doesn't need it either. And so the real work lies in the gray zone, and that's what we're trying to help with here, Zach.
Zachary Klaassen: Yeah, great summary. And we'll tag your paper, the citation to our discussion. Thanks for joining us, Nate.
Nathan Perlis: Appreciate it. Thanks for highlighting this paper, and be well.
Zachary Klaassen: Thanks.