Read the Full Video Transcript

Alicia Morgans: Hi. I'm so excited to be here today with Professor Tamara Lotan, who is joining me from Johns Hopkins University. Thank you so much for being here.

Tamara Lotan: Great. Thanks for having me.

Alicia Morgans: Wonderful. Tamara, you are a pathologist. And we don't get to interview a lot of pathologists on UroToday, so I'm really excited to talk with you and really about the overlap of pathology and urologic oncology, and specifically how this has evolved in a really digitized and even AI type of world. Can you tell me a little bit about the progress that's been made in pathology and what we should be on the lookout for now as we're thinking about modern-day pathology?

Tamara Lotan: Yeah, these are great questions. I always start by telling people pathology has really not changed for the last hundred years or so. The way we practice, we use the same kind of microscopes, small updates. We still take the tissue and fix it in fixative and process it and cut it onto glass slides. Been doing this exactly the same way for a hundred years. So I think that the rise of digital pathology is really exciting, and that really is a step where we then take the glass slide and scan it so that we have a digital representation of that that we can look at on a computer screen. And then most importantly, use it now to apply all kinds of algorithms, especially machine learning or artificial intelligence algorithms, to try to augment the diagnoses that until now we've just been making by eye through the microscope. We can now make with some additional machine learning techniques that really, I think, augment our ability to make diagnoses and also to predict how patients are going to do in the future.

Alicia Morgans: Well, that's certainly really exciting. I wonder if you can walk us through some use cases. What are situations where we might actually see this?

Tamara Lotan: Yeah. Prostate cancer has really been a test case for the use of digital pathology from the beginning, I think because it's a very high volume practice, looking at prostate biopsies, for example. And a lot of what we do with prostate biopsies is semi-quantitative where we identify the cancer and then grade it by deciding what percent pattern 3, 4, 5 we think is in the tumor. And we know that humans don't do that very well by eye visually. We can only get to a certain accuracy visually, whereas a machine learning type algorithm where we can teach the machine to recognize these different patterns that the tumor cells are making and then quantify the relative proportions of those patterns can do that much more accurately.

We know when humans compare, if I grade a prostate cancer case and I compare to my colleague, probably only agreement, we say CAPA values is how we typically do the metric of agreement. And it's usually between 0.4 and 0.6, which is at most moderate agreement. And most clinicians, I think, have had that experience where they see a case that's seen at their institution. Outside institution, you get slight disagreements. So the hope, I think, is that these algorithms will help us for diagnosis, but then maybe most importantly for grading so that we can really standardize grading. Institutions that don't have urologic pathologists who have had a lot of extra training and maybe have higher inter-observer agreement in terms of grading are not available at all institutions, and certainly not internationally in many areas of the world. So if you had an algorithm that can do this more accurately and represent what a urologic pathologist would interpret on the slide, that would really equalize access to care.

So I think that's very exciting. And also just ensure a consistency in terms of grading, which we know is really the most important prognostic parameter in prostate cancer.

Alicia Morgans: I couldn't agree more. And it's so critical that I tell second opinions that come to see me in my clinic that perhaps even more than what I say is that second opinion by the pathology group, because understanding the risk based on the Gleason score, of course, is the way that we really need to personalize treatment and make those decisions in the localized disease setting. So what's another use case? Where else might we see the use of this?

Tamara Lotan: I think one exciting area in prostate is thinking about how we can go beyond Gleason grading. As you mentioned, Gleason grading is such a powerful prognostic parameter, but it actually was developed in the '60s based on I think only 300 or so cases in one VA hospital. So it's amazing that it performs as well as it does, but there's no question we can probably do even better than Gleason grading in terms of coming up with a totally novel system that predicts patient outcomes.

So I think another exciting use case is to think about training algorithms on large, large data sets where we know exactly how each patient did. And we can ask the computer to look at this totally agnostic to any human grading system and come up with features that are associated with prognosis and a totally novel prognostic prediction algorithm. And I think those kind of data sets may even improve on Gleason grading when we can do that, not just with 300 cases like Dr. Gleason did, but now with thousands and thousands of cases, much more diverse data sets, for example as well.

Alicia Morgans: Is something like that in the works right now, is that something that we should be on the lookout from Lotan, et al?

Tamara Lotan: Yeah. We and many other groups, I think, are working on developing exactly these algorithms. One challenge has been to just get enough cases that have these very high quality clinical annotations and also have whole slide images available. But yes, many groups are working on putting together those data sets and maybe even sharing them across institutions. So we really have multi-institutional data sets, which are especially critical in digital pathology.

Alicia Morgans: That'll be really exciting because not only will that help, I think, make the diagnostic process more equitable. It will be fast because you can run it through those kinds of tests in a really quick manner and get an answer to a patient, to a clinician to really get moving on a treatment plan.

Tamara Lotan: Yeah, a hundred percent. It also doesn't exhaust any tissue. This is using just the original diagnostic H&E or hematoxylin, eosin stain. So you don't have to worry about using tissue that you may then want to preserve for future genomic assays and things like that.

Alicia Morgans: Great. So what's another use case as we continue to pepper you with these questions?

Tamara Lotan: Yeah, sure. I think another great use case is a long prediction of response to therapy. So there already are some companies working in this space and academic groups also studying large data sets from clinical trials, for example, to try to predict. Does this patient need hormonal therapy in addition to radiation, for example, or perhaps they would respond just to radiation alone? So I think these, we're going to see more and more predictive biomarkers coming out from AI or deep learning algorithms that are using just the diagnostic pathology images, and that'll be very exciting.

Another area I think that is definitely growing is thinking about how we can maybe triage patients for downstream sequencing. We know, although we recommend patients have germline and somatic sequencing if they're high risk, in some cases, not everyone is getting that sequencing even nationally, certainly not internationally. So if we had, for example, machine learning algorithms that can look at these diagnostic tissue samples and say, "This patient has a 90% risk of having," for example, "a BRCA2 mutation," then we could really flag those cases that we think need sequencing especially and make sure that that happens in those cases.

Alicia Morgans: That would be incredibly powerful because I think if clinicians, particularly those busy clinicians who are maybe in a community setting where they see so many patients in a day from all different types of cancer and don't necessarily know the specifics of this test or that test for their prostate cancer patients, that would be just a wonderful way to raise a flag and say, "Make sure you get germline and somatic testing for this patient." That's great.

Tamara Lotan: Yeah, a hundred percent. I think that's very exciting.

Alicia Morgans: All of this excitement I think needs to be couched with our understanding of the limitations and the challenges that the field faces. And what are those?

Tamara Lotan: So probably the biggest challenge in terms of uptake, I think, in most pathology labs is the cost. So digital pathology is really not budget-neutral. It's adding cost at least right now because we still have to prepare the tissue in exactly the same way we would to just look at it using a glass slide under the microscope. But then we need an additional step where we buy often very expensive scanning, digital scanner machines often costing several hundreds of thousands of dollars. And if you're going to do the throughput of a very large pathology lab, you might spend millions and millions of dollars on this equipment. And then even more than that, that's a fixed cost, but even more than that is the cost of storing these images. So they're huge images actually compared to radiology images. A typical digital pathology image for a single slide might be three gigabytes or something like that. So you end up terabytes, petabytes of data and the expense of that, which just accumulates over the years. In labs like ours, which are producing a million slides a year, it becomes astronomical.

So I think a lot of groups are thinking about how to store images similar to our radiology colleagues where you have some that are immediately accessible and some that are in a colder kind of storage that reduces cost, for example, and hoping that the cost for storage goes down over time. But I think that what's really going to drive uptake is if we have these compelling use cases. So if pathologists are saying, "I really want a digital pathology practice in my group because I feel like that will make my Gleason grading more reproducible," if a clinician like yourself says, "I want you to put a test online that predicts the outcome of my patient better than grading alone," that hopefully will drive investment by big health systems in this technology.

Alicia Morgans: I agree. Once certain groups are doing it, then other groups will follow, and so on down the line. This is really, really important. So if you have to look back and think about digital pathology, these AI algorithms, machine learning, what would your message be to listeners as they try to think more deeply about pathology and the way that it intersects with our care of patients with GU malignancies?

Tamara Lotan: I think this is an incredibly exciting time. I think we're really at the precipice of this revolution in how we practice pathology, and we're only going to become more accurate, more reproducible, and more equitable across all kinds of care settings. So I think that's a super exciting part of being a pathologist right now. Many pathologists or many, I would say, non-pathologists say to pathologists, "Oh, you must be scared. You won't have a job in the future. You're going to be replaced by machines." But I think pathologists really look forward to this because I think this will take away the most menial parts of our job where we're just screening things and replace it with a much more efficient process and really allow us to focus on the things that are most exciting and that require all of that medical knowledge that we've accrued over the years to apply to our patient samples. So we're really looking forward, I think, to the future, and I think it's a very exciting time for the field.

Alicia Morgans: Well, I could not agree more, and I really look forward to seeing the bright future that is the new wave of pathology in GU oncology but also of course, I'm sure across the spectrum of cancers and perhaps even beyond. And I thank you so much for sharing your knowledge today, for your time, and for your expertise.

Tamara Lotan: Thanks so much for having me.