2. UroToday Home
  3. Video Lectures
  4. PCF 2025

AI Applications in Prostate Cancer Pathology and Prognostics - Tamara Lotan

November 20, 2025

Tamara Lotan discusses artificial intelligence applications in prostate cancer pathology. Dr. Lotan describes three main categories of AI tools: diagnostic algorithms for tumor annotation, prognostic models for outcome prediction, and predictive tools for therapy response. She emphasizes that pathology-based AI models for Gleason grading are ready for routine clinical use, addressing inter-observer variability and providing quantitative, reproducible results. The conversation highlights AI's potential to democratize expert-level pathology globally by standardizing grading, with the primary barrier being slide digitization rather than algorithm development. Dr. Lotan envisions increasingly multimodal models integrating clinical parameters, genomic data, and spatial omics technologies over the next two to five years. The discussion emphasizes the importance of federated learning approaches and population-based validation studies to ensure AI tools perform effectively across diverse patient populations internationally.

Biographies:

Tamara Lotan, MD, Molecular Pathologist, Professor of Pathology, Oncology and Urology, Johns Hopkins University School of Medicine, Baltimore, MD

Zachary Klaassen, MD, MSc, Urologic Oncologist, Assistant Professor of Surgery/Urology at the Medical College of Georgia at Augusta University, Wellstar MCG, Georgia Cancer Center, Augusta, GA



Related Content:

ASTRO 2025: Digital Pathology-based Multimodal Artificial Intelligence Biomarker Models in Oligometastatic Castration-sensitive Prostate Cancer

ASCO GU 2025: Comparison of Digital Pathology AI Models, Genomics Classifier, and Clinical Variables in Predicting Progression Free Survival in TCGA Prostate Data Set

AI's Role in Prostate Cancer Diagnosis, Imaging, and Pathology - Matthew Cooperberg

Predicting Prostate Cancer Molecular Subtype with Deep Learning on Histopathologic Images - Tamara Lotan
Read the Full Video Transcript

Zachary Klaassen: Hi, my name is Zach Klaassen, urologic oncologist at the Georgia Cancer Center in Augusta, Georgia. We are out at UroToday at the Prostate Cancer Scientific Retreat in Carlsbad, California. I'm delighted to be joined by Dr. Tamara Lotan, who is a pathologist at Johns Hopkins University. Tamara, thanks for joining us on UroToday.

Tamara Lotan: It's great to be here. Thanks for having me.

Zachary Klaassen: Absolutely. And so you had a session at the Prostate Cancer Scientific Retreat looking at artificial intelligence, which has just exploded over the last several years. And so it's important to have these kind of sessions at retreats like this. So just maybe give us a high level overview of what that session looked like.

Tamara Lotan: Yeah. Yeah, definitely explosion in different AI or artificial intelligence-based tools for prostate cancer. Really leveraging, I think, huge data sets that we've developed over the last decades as we've treated patients and had more experience with them. So we have clinical data sets with patient outcomes and clinical pathologic patient parameters. We have imaging data sets around radiologic imaging and pathologic slide imaging to look at tumors, large genomic data sets now for the tumors themselves or germline genomic data from the patients. And the idea is we really need these AI tools to start to integrate across these. And the session was wonderful in the sense that it kind of introduced what I think of the three main AI tools that we have at our disposal now. We have diagnostic tools. Those are mostly in prostate for pathologists.

Zachary Klaassen: Sure.

Tamara Lotan: And I could talk a little bit more about that. And then we also have prognostic tools that help us better estimate how patients are going to do and then predictive tools in terms of response to specific therapy. And the idea is to kind of create AI models either in one domain with one type of data or having multimodal data inputs, which is I think really where we're going in the future to work on those models.

Zachary Klaassen: It's so interesting because it seems like even just a couple years ago, as I thought about our AI was like pathology, radiology makes a lot of sense, but we're seeing it in so many things. The predictive biomarkers, it's just crazy. I know you gave a talk also about localized prostate cancer, probably from a pathology perspective. Just give our listeners an overview of what the highlights were of that talk.

Tamara Lotan: Yeah. So within pathology based AI models, which is a huge growing field right now, prostate cancer has really been kind of at the forefront. I think because we already started with sort of a semi-quantitative grading system, so it looked like a very easy place to introduce computer vision kind of models to look at the tissue and estimate, diagnose a tumor, annotate the tumor, and then estimate the different contributions of various Gleason patterns.

Zachary Klaassen: Sure.

Tamara Lotan: And so, I think at this point we really have models that are ready for primetime, in fact being used by many of my colleagues on a daily basis that can annotate slides for you in advance to tell you which are likely to have tumor and which are not, so that you can prioritize specimens as you look at them, and of course give us very quantitative and reproducible data on Gleason grading, because Gleason grading is somewhat subjective and we've had a lot of issues around inter-observer variability. And then we're sort of introducing, on top of these, maybe more bread and butter kind of workhorse pathology type algorithms, algorithms to move beyond Gleason grading and really estimate prognosis just directly from the histology of the tumor. And that's a very exciting area.

Zachary Klaassen: Absolutely. And I think one thing you said I wanted to spin off a little bit, and you mentioned just the standardization and some of the Gleason grading. From a global perspective, I mean, you guys are at the epicenter of Johns Hopkins, even in this country where blessed with GU pathologists. Say in a developing country, where do you see AI could maybe fill in that gap of not having the training of a GU pathologist?

Tamara Lotan: Yeah, that's a fantastic question. I think not even a developed country, but even in parts of the US-

Zachary Klaassen: Sure. Yeah, fair enough.

Tamara Lotan: ...Where they don't have as many specialists. Yeah, I think this is really a game changer in terms of standardizing grading, giving everyone ... Imagine that the whole world could have access to the types of grading that we see at these sort of tertiary care institutions like Hopkins, which we have 30,000 prostatectomies-

Zachary Klaassen: Wow.

Tamara Lotan: ...In our archives, so you can imagine the level of just experience from seeing that amount of volume. If we can make that available, that standard of grading, to the whole world, which is completely feasible at this point even.

Zachary Klaassen: Sure. Yeah.

Tamara Lotan: The real barrier I think is getting to the point of digitizing the slides.

Zachary Klaassen: Yes.

Tamara Lotan: Right? That's where we lag behind radiology, where the images come off the scanner-

Zachary Klaassen: It's all there.

Tamara Lotan: ...Already digital, we have this extra step where we have to scan them and that's resource and time intensive.

Zachary Klaassen: Yeah, great point.

Tamara Lotan: But even now there are newer, I think on the horizon, imaging techniques to avoid creating even a glass slide and just imaging small pieces of tumor in 3D. So we may see that in the next decade or so.

Zachary Klaassen: Absolutely. And I think too, I mean, the ability for these models to get better with more data, I mean, that's also pretty cool too because you just keep feeding it more, it keeps getting more precise. And I think, from a global standpoint, would be just incredible. Where do you see the field going over the next say two to five years? I know it's hard to predict, but that's the fun part of having these conversations is sort of figured out where it might go, but what are your thoughts on that?

Tamara Lotan: Yeah, 100%. I think we'll see increasingly multimodal models. So right now, some of the talks focus on just pathology based models, some models integrate some clinical parameters, but I think we're going to see increasingly multimodal models, where we now integrate maybe even raw genomic data from the tumor or large germline genomic datasets from the tumor, probably increasing new technologies that are even more data intensive, like spatial omics kind of technologies where we not only just image the cells with an H&E, but now we actually know something about the transcriptional state of all those cells and how they relate to one another. Clearly, AI models will be needed for that scale of data and that may reach patients. So I think those are two areas. I think we'll see also additional, maybe more federated learning where we allow institutions to essentially share data by having the model go to the data instead of the data come to the model, which requires a lot of complicated agreements.

Zachary Klaassen: Great point.

Tamara Lotan: And hopefully also more sort of population based studies where we hopefully kind of learned from some of the mistakes in the early genomics era where we studied just one population now and missed a lot of information about other populations.

Zachary Klaassen: Sure.

Tamara Lotan: We really need to test these in populations all over the world, essentially.

Zachary Klaassen: Yeah. Great summary. I think it's going to be super exciting. We've seen, like we started off saying, there's been an explosion of data, seems like it's just going to continue for the foreseeable future. So great conversation. Any take home messages, concluding statements, anything we haven't touched on yet?

Tamara Lotan: No, I think just that prostate cancer is really kind of at the vanguard of this field and it's going to be really exciting to see what we can accomplish across all these domains. And hopefully, eventually we'll have models that we've talked now really about models that clinicians will use maybe to predict patient outcomes, but there are also going to be lots of models that talk directly to patients with the appropriate caveats and so forth. But I think those will be really important for patients moving forward.

Zachary Klaassen: Well said. Tamara, thanks so much for joining us on UroToday.

Tamara Lotan: Wonderful. Thanks for having me.