One of the Deadliest Cancers in America May Have Met Its Match

Today, one of the top scientific breakthroughs of the year is in the War on Cancer. Pencratic cancer is one of the deadliest cancers we know. It killed more than 50,000 Americans last year, most of them within 12 months of the diagnosis. It is hard to detect, aggressive once it shows itself, and brutally indifferent to standard chemotherapy. By the time most patients learn they have it, it is already too late. Millions of people around the world have, like me, lost parents, siblings, or children to this disease. There are two reasons why pencratic cancer is evaded modern science. The first is genetic. Most pencratic cancers are driven by mutation in a gene called K-RAS, and for 40 years K-RAS has been considered "undruggable," a smooth slippery target that no doctor's molecule can grab onto. The second reason is our immune system. Pencratic cancers carry relatively few mutations, which means they don't waive many red flags for RT sales to go out and kill. They grow quietly in the dark for years. In just the last few weeks, we've gotten remarkable news on both of these fronts. First, a company called Revolution Medicines reported results from a small trial of a new drug that targets that genetic mutation directly. In patients with late stage pencratic cancer, the drug shrank tumors in nearly half of those treated. Next week, the FDA gave RevMed a green light to expand access to that medicine. Second, a team at Memorialstone Kettering and the German company Beyond Tech reported follow-up data on something even stranger and more remarkable. A personalized mRNA vaccine built using the same technology behind the COVID shots that teaches our immune system to recognize a patient's own cancer. Here is the lead researcher on that project for Node Balachandrin explaining to me, last year on this podcast, exactly how that vaccine works. In this trial, we did surgery here on patients at Sloan Kettering in New York. Within 72 hours, we shipped the tumors to colleagues in Germany who then do genetic analysis of the tumor, create a bisput vaccine, ship it back to us, and then we treat patients here in New York and then watch how the patients do and perform deep scientific analysis in them. So, we had vaccinated 16 patients in this trial. In eight of the 16 patients, these vaccines made lots of T cells. We call these eight patients responders. And in 2023, when we had looked at an average year and a half follow-up, we had reported that among the eight responders, none of the responders had seen their pancreatic cancers return after surgery. And in contrast, eight of the non-responders, six of eight of these non-responders had seen their cancers return after surgery. Cancer's power lives in its camouflage. The immune system is often described as a kind of military operation with our T cells acting as snipers, hunting down foreign invaders. But cancer kills so many of us because it doesn't look foreign. It looks like us. In his book, The Song of the Cell, Siddhartha Mukherjee writes that the proteins cancer cells make are, with very few exceptions, the same proteins that normal cells make. Cancer simply distorts their function and hijacks the cell toward malignant growth. He calls this, quote, the oncologists nemesis, end quote. The disease is kinship to the self and its invisibility. To attack a cancer, he writes, you first have to make it visible to the immune system. Pancreatic cancer is, in this way, the invisible emperor of all maladies. For decades, scientists wondered whether the immune system could be taught to see it at all. But the Sloan Kettering Beyond Tech vaccine is the first real proof that the answer might be yes. And Balachandran believes the same approach could eventually be turned on to other cancers, too. All of this brings us to today's episode and the third piece of this puzzle. Pancreatic cancer is so deadly in part because it is almost always caught too late for effective treatment. So what we should really hope for, beyond first line therapies, beyond secondary vaccines, is a way to see pancreatic cancer on a typical screen, a typical CT scan of a gut. And we might just have the first indication that we can do that. Last week, the Mayo Clinic reported that their new AI tool could help specialists detect pancreatic cancer up to three years before a typical diagnosis. So you put it all together, a drug that targets the gene driving most pancreatic cancers, a vaccine that teaches the immune system to recognize the invisible. And third, an AI that can spot the cancer in scans years before any human radiologist would catch it. For the first time in a long time, the deadliest cancer in America looks like something we might actually be able to fight and even to cure. Today's guest is Dr. Ajit Goenkup, a radiologist at the Mayo Clinic who studies AI imaging and was the lead author of this new study. We talk about his research. Why AI seems so good at finding cancer? Whether this news is, as some AI stories turn out to be, too good to be true? And what medicine might look like in a world where artificial intelligence can read our bodies better than human doctors can? I'm Derek Thompson. This is Plane English. Dr. Ajit Goenkup. Welcome to the show. Thank you for having me. Why do you get us started by telling me who are you and what do you do? Yeah, I'm a consultant radiologist and a professor of radiology here at Mayo Clinic and Rochester, Minnesota. And what I do is that I try to solve the complex problems for the patients who come to Mayo and otherwise. So, you know, that's the 100,000 view of what we do. We can go into the nitty gritties as we speak along. Great. I mean, the first thing I want to talk to you about today is this new remarkable study of AI radiology and pancreatic cancer, which was just published in the journal Gut headline finding. Tell me what you found. Well, the headline finding is that 85% of the patients who develop pancreas cancer, they hear the word pancreas cancer at a stage where it is too late for them to do anything about it. What we are trying to do here is that we are trying to flip that equation. And the way we are trying to flip that equation is that we are trying to find those signals, those mathematical signals from the images that can tell us well before time whether or not we can cure a particular patient with pancreas cancer. So, AI is just a tool that we are utilizing to solve that problem. Our goal is early detection and what this study shows is that it is eminently possible. There is certainly a lot more work that we have to do. But I think we are at the base camp now and it's a matter of time before we start climbing the Everest. So, that's the headline finding. I would really like to understand what you did, like how you found these patients, how you found these scans. And in particular, how you allowed AI to read the scans without cheating because there's been some accusations that these AI assisted imaging readings involve the AI like going online looking up the patient's medical records and saying, "Oh, well, that's pancreatic cancer." Or, "Oh, nope, no pancreatic cancer." And so, you're actually not testing whether the AI can see, well, at all. You're just looking up whether we're studying whether the AI is accessing the internet. So, how did this study work and how did you make sure that the AI was not cheating? Yeah, that's a great question. And you're absolutely right. That's the fear that all of us dread is that we don't want to come out of the study that does that. So, number one, you know, the great thing about being at a place like Mayo Clinic is that we get patient referrals from all over the world. So, we had about 5,500 patients in our archives. We went back in time to find out who had a CT scan that was done for an unrelated indication three months to three years prior to the diagnosis. Now, you know, you mentioned about how we ensure that the AI was not cheating. So, one of the ways we could have done that or we could have allowed the AI to cheat is that on those CT scans, we could have taken those camps where there was a cancer present. You could see but it was missed. We didn't let that happen. The way we didn't let that happen is that two things. We had few team members of ours radiologists who looked at each and every one of that CT scan to confirm that there is no cancer present. So that is one way you did that. Second thing we did is that we took the right controls which means that patients who did not have cancer. We made sure that they were demographically and time wise comparable to what we had in our test set. So that way we ensured that the AI did not have opportunity to kind of in a way learn the noise. Oh yeah, that CT scan which is a control looks a little different. So I'm going to use that information to make my prediction. So that's another way. Those are the two big ways that you know one can cheat and we were very careful and deliberate about it. As I was reading this study, I saw that in scans obtained 18 months before diagnosis the AI radiology was twice as sensitive as radiologists and in scans obtained more than 24 months before a diagnosis was made, it was three times more sensitive. So to my untrained not a doctor ear, that sounds like AI assisted radiology in this study was roughly three times better at finding pancreatic cancer than the expert radiologist. Is that a fair conclusion to draw? The short answer is that yes, but you know when it comes to clinical practice because eventually we are not doing this for presentations or for media attention. I mean, those things are great. It is great that we are talking about a disease that is long overdue. But if you look into clinical practice beyond just its ability to find cancer, what is also important is ability to tell somebody that no, you don't have cancer, which is a specificity. So all of those things have to be taken into account. So the short answer is absolutely right, but there are more nuances to it, which I'm happy to go into details with you. Yeah, and one thing that I always wonder about when it comes to diagnostics and the world in which we're going to get more and more tools for figuring out the ways in which we are sick or might become sick are these two words of sensitivity versus specificity, right? Sensitivity meaning can you find the positives that are there and specificity meaning are you ruling out the negatives, right? Because what I don't want is to like get a full body MRI every single year. It finds these like 10 little cysts in my body that are never going to become anything. And then tells me, Derek, you have 10 possible developing cancers in your lung, in your gut, in your leg. Okay, well, that's going to ruin my life. It's going to ruin my life emotionally, financially. I might go through a bunch of tests. So it's going to ruin me physically. We don't want that. How well was this test at weeding out the positives from the negatives? Yeah, so that's an excellent question. So, you know, to add to the complexity of the terminologies here, in addition to sensitivity and specificity, there is a term called accuracy that in a way integrates all of those concepts into one metric. And that metric in this case was about 0.84, 0.85. So, which means that 85% so when you take both those things into account, it was about 85% overall. Now, more than good. I think that is pretty good because, you know, the reason why it's important is because you have to compare AI to what is out there, right? You cannot compare AI or any tool for that matter to perfection because there is no perfection. And right now, if you can see what it was, you know, you have the answer. All of these CT scans were called negative. So in this case, your bar was pretty low to begin with. So when you take a bar like that and you go to 85% then I think that some monumental accomplishment on the part of this tool. But having said that, not just sensitivity and specificity, what I think most matters in the clinical arena, in the practice to address a problem that you mentioned is that what is the pre-test likelihood of somebody developing this particular disease? So to give you an example, you know, since the time we published this particular thing, we've had, you know, dozens, hundreds of queries from concern, you know, patients and family members all over the world. And they are absolutely justified in asking those questions. But here's the thing is that you can have a test that is 99.99% sensitive, 99.99% specific. But if you apply that test to a patient population where the pre-test likelihood of them developing that particular disease is very low, you will still get hundreds and thousands of false positives. And you know, to give you the context, the whole body MRI that you mentioned, is in a way, you know, kind of arch-tapical of that particular problem. Someone like you who's young and healthy otherwise should not be getting that test because you will inevitably not show anything, but you will end up with a lot of these incidental findings that are going to wreck havoc with your life because, you know, as an individual, you are concerned. You may be reading the statistics, but you might say, but what if I am one of those one in 100 where the cyst goes on to develop cancer? So the message here is that this can be applied only in the right patient population, which in this case would be those individuals over the age of 50 who have got certain risk factors that puts them at a risk of pancreas cancer that is high enough to justify an early detection paradigm like this. Well, you perfectly anticipate my next question, which is, let's assume that this technology works. And I do still have questions about making sure that it does work, but let's assume the technology works. How do you apply it in the real world? So you can tell a story here, but I'll tell my own story, right? My mother passed away at pancreatic cancer when she was 63. In genetic testing, I have seen that those tests have come back showing that I have an elevated risk for pancreatitis, so not necessarily pancreatic cancer, but some other bad stuff happening to the pancreas. I'm 39 years old, turning 40 in a few weeks. When would I take a test like this, an AI enabled CAT scan at a clinic like Mayo, when would someone like me be taking a screening like this? Because it's not that helpful if we're just applying it to the entire population. Absolutely. So no test can be applied to the global population. And the reason for that is because if you look at pancreas cancer specifically, there are a few things we know about it for a fact. Number one is that it is the most deadly cancer we know of. By 2030, it will be the number two cause of cancer deaths in the United States. 64,000 Americans will be diagnosed with it every year, and almost an equal number will unfortunately succumb to that disease. So it is, it is a deadly cancer, but again, it's only 64,000. Now in isolation that number is a lot, but when you compare it to something like prostate cancer, lung cancer, breast cancer, colon cancer, I mean, those are far more frequent compared to pancreas cancer. So what you have to do is that you have to take your risk quite area and stack them up so that your risk becomes so high that even that incidental findings that you mentioned are justified. Even the cost is justified. Even the small amount of radiation dose from the CT scan is fully justified. So that is what we are already doing as part of a prospective clinical trial. We call it AI paste, which stands for AI augmented pancreas cancer early detection. And the name is actually quite apt. You know, we are really running against time in this particular cancer. And in that prospective trial, which is right now on clinical trials.gov, we are taking individuals above the age of 50, which answers one of your questions, who have two very well established risk factors. One is what we call a new onset diabetes. So this is not your common garden variety of diabetes, which many people have. This is something which is very characteristic in terms of its calisonic parameters. I mean, what happens to your blood glucose, how quickly does it rise? Second thing that it does is that it takes those people with new onset diabetes and further stacks them up in terms of risk factors by looking at what we call end-packed score. So in summary, we take those individuals who have new onset diabetes, who have a very high what we call an end-packed score, which then, you know, we discussed earlier about pre-test likelihood. So their pre-test likelihood or developing pancreas cancer is high enough in whom such a modality would be justified. So that is essentially what we are trying to do as part of clinical trial is to see that in the real world, how does this technology do when it is better tested? So we have that trial that is running on. It will take about three to five years. And the reason it takes three to five years is got nothing to do with AI. It has everything got to do with the design of any kind of a study because when you take those individuals, when you apply AI to their CT scans, you have to follow them up for the next three to five years to see if they go on to develop cancer. And that's only the way you would know whether or not the prediction you got from AI is right or wrong. So that is, you know, I'm assuming a question that would likely be on the top of the mind of many of your viewers as well as a few. Yeah, and please reiterate this because this seems very important. I think a lot of people who are interested in artificial intelligence believe or hope or predict that the effect it's going to have on science is almost immediate, that almost immediately because of superintelligence or advances made and finding certain kinds of proteins, we're going to solve cancer imminently. We're going to solve untreated diseases imminently. And if we slow down a little bit, it sounds like what you're saying is that in order to do good science, you need to make sure that the interventions are having the right effect on the patient population. And that simply takes three to five years so that you can come back and say, okay, what's the difference between the group that we intervened in and some control group that we can compare it to? That just by definition takes years. And so while this is incredibly promising, we also shouldn't expect that it will will necessarily change standard of care in the next few months. Is that a fair recapitulation? - Yeah, I mean, here's the main difference between Silicon Valley, which has done pioneering work in advancing AI, where the motto is that move fast and break things, right? In healthcare, we live by the motto of, first, do know harm. So what we want to make sure is that we are deliberate, strategic, and ensuring that none of our patients get harmed in the process, where we are taking these technologies and introducing them in clinical care. Now, it's very easy for any kind of a healthcare system to say that, all right, you know, this works great. Let's just roll it out in a clinical practice, we'll deal with whatever comes on. But no, that's not what we do at a places like Mayo Clinic. So that is the reason why we are taking this through a systematic rigorous process. Because, you know, as I said, our goal is not AI, our goal is early detection. So if AI happens to be that means that takes it fair, so be it, but if it doesn't, we'll switch. - Dr. Wood, I would say to that is, there are at least two ways to do harm. One way you can do harm is by using a technology or a method that isn't proven and hurting patients in the process. But there's another way to do harm. And that might be withholding a technology or therapy that might be promising from a patient who needs it. So, you know, if a patient's coming in to the Mayo Clinic or some other clinic, and they're complaining of mysterious stomach or back pains, and they're in their fifties, and they have a family history of pancreatic cancer, let's say. And maybe there's even a genetic predisposition to then developing something like pancreatic cancer. - Well, wouldn't we want the radiologists looking at those scans to be assisted with artificial intelligence? Given the enormous amount of information we're beginning to get about how AI helps radiologists see, mathematically as you put it, what the human eye alone cannot see. I mean, shouldn't we still be interested in using AI in radiology relatively aggressively considering everything we're learning? - Excellent point. I mean, you've touched a very important point. So here's what I'll say about that. So first and foremost, if somebody has risk factors or a pancreas cancer, and they already have symptoms, believe me, you don't need AI to be able to diagnose pancreas cancer in them. Because if they have symptoms, then often, almost always, it's too late. - Right. In that case, oftentimes a medical student can look at a CT scan and tell you if it's right or wrong. So that is the challenge with pancreas cancer is that it does not scream, it whispers. And so this is what we are trying to do is that we are trying to amplify that whisper. And so therefore, you know, so one of the ways we could kind of build upon that hype of AI is by saying, promising something like that. Because, you know, if we do that, then every time it will come back as true positive. But you don't need that. You have to provide incremental utility above and beyond what your standard of care is. So that's the nuanced view point against that argument. - Right. Is there another kind of scientific breakthrough that would help limit the patient population that you're using this technology for so that there's a higher batting average, so to speak, for AI assisted radiology. So for example, if we did develop a blood test that was highly accurate at predicting susceptibility to pancreatic cancer, then something like that could potentially be paired with a predictive AI radiology screening so that together, it's not like you're looking at 300 million Americans every year for pancreatic cancer. You're looking at the right, quote unquote, 500,000 people a year in order to detect to detect whether or not there's something going on with their gut. What is the missing technology, the missing science, the missing piece here that would help you narrow this population pool? - Yeah, absolutely. So what you have laid out is exactly the paradigm that we are building. So one of the challenges in having a blood test like that is let me explain to you. So you know, initially you mentioned about how we did this. We went back in time of those patients who had pancreatic cancer to find out if they had an incidental CT scan, somewhere in the patient chart, right? Now here's a problem, you know, nobody like that has got a blood bank somewhere in the world in anticipation that I would develop cancer, right? You know, who does that? It's not, it's just not a thing. So the problem is that we don't have pre-diagnostic blood samples unlike the pre-diagnostic CTs that we were lucky to have at the Mayo Clinic. So here is how we are tackling that. As part of this clinical trial, we are not just doing AIS state CT. We are also collecting blood. So which means that we are banking that blood because we do know that whatever we can do to enrich that risk to a level where we know that, all right, this is an emergency. It needs to be, everything needs to be thrown at it to be able to find that cancer. So we are collecting that blood, we are banking that blood, which will be the pre-diagnostic specimen. So it creates a, you know, pressure for us to be able to do that biomarker discovery, which is very critical as you are rightly pointed out to be able to make the AIS succeed or whatever a stool that we may use to succeed in that context. So we are doing that. Unfortunately, right now there is no test out there that can do it. Not only that, we are doing a few other things. One thing is that we are not waiting for that three to five year period. We are doing what we call as in silico clinical trials. So what happens, you know, one of the advantages of AIS in general, not in this case, is that you can do modeling and simulations to a very high precision. So we're doing exactly like that. So the questions we're trying to answer is that, okay, if you have a pathway, whereby you take a particular group of patients. Now, one of the things you can decide is that, do you want to have a blood test right there? Do you want to have a blood test after the CT scan? Or do you want to have altogether? But everything has a cost, you know? You know, so that's what we are doing is that we are building those simulations. We are finding out what is the incremental benefit in terms of information achieved, in terms of cost effectiveness, in terms of outcomes. So we are doing that already. And I can tell you on that front, those results will be out before the end of the year. And why those results are important is that they provide a pathway for many other stakeholders. You know, we are not the only ones. There are regulatory agencies, there are policy makers, there are people who decide whether or not these things should be reimbursed. And all of that, they work based on data. So it is our job to be able to create that data rigorously and transparently and put it out there for them to decide. - Given all the fronts that we're seeing progress on with pancreatic cancer right now, you've got Revmed and its K-RAS inhibitor drug. You've got the Balashandran lab and its secondary vaccine. You're looking at AI-assisted CT scans. Can you paint me a picture of what treating pancreatic cancer could look like five to 10 years from now if everything goes right? Like what kind of a world might we be walking into? - Yeah, so that's excellent question. We already have looked at that scenario and some of the people that are coming out from our group have already painted that scenario for people to know. So the short answer is that right now, cure is not part of the vocabulary, typically when a patient walks into the clinic with the diagnosis of pancreatic cancer. But in that future, cure will be the only thing that will be part of the vocabulary and here is the reason why. So you mentioned about some of the recent development. One of the recent developments has been where they have taken pre-clinical models of pancreatic cancer. So they have taken mouse models. And what they have found out is that in those mouse models, if you give them K-RAS inhibitors, before the tumor develops, they can actually preclude the possibility of the tumor ever developing. So in other words, here's the implication for what we are doing. If we can pick it up at a stage where we have shown it can be done, which is at a stage where it is visually occult, which means that for all practical purposes, it hasn't developed. So if you can pick it up at that stage, we take those drugs and we give to these individuals where we are certain that it will develop if we don't do anything, then there is no reason for it to develop at all. So we have other cancers where we have made remarkable progress so for the decade, you know, lymphoma for instance. You know, some of the testicular cancers for that matter, you know, where people have had that and then at the end of the day it becomes like common cold. Oh yeah, I had it about 10 years ago. They treated me, it's all gone. You know, once in a while I may get a surveillance scan, that's about it, I'm living my life. So that's the vision, that's the outline. we want to deliver to our patients and you know you you're absolutely right that all of these promising developments that are happening in the domain of pancreas cancer are fortunately coming together in a way that provides us a picture of how it could be and that's why it is so important that we talk about this disease because in it will take a lot more than what we are doing right now it will require more funding it will require more resources it will require more like-minded people working together to be able to get there. Dr. I just want to point out that I listed sort of three different fronts that are advancing in the war and pancreatic cancer. This new drug from Revmed that is taking on the genetic mutation of people who have pancreatic cancer. Number two, this secondary vaccine from the Balachanjren lab Memorialstone Kettering. And number three, your AI imaging. For the folks listening and watching at home I want to be clear that your answer included a fourth front to open that we're just beginning to work on in mice a different kind of inhibitor a different kind of almost preventive cancer medicine that like finds or treats this genetic mutation before it starts to spin out of control. Is that right that you're talking about a fourth technology that if it's introduced the picture okay now we're really cooking with gas we can start to use the new C word which is cure. Yeah absolutely and and you know there's a very specific term that exists for that paradigm and that is called pre-clinical interception. So the reason it is pre-clinical is because they haven't developed symptoms. Interception is because you can't really use a word cure because they don't have the disease. You're intercepting it before it develops a disease and here is what I want to drive home. All of those developments underscore the importance of detecting it at a stage where we are trying to detect it because if you can't detect it how do you know you have to give that drug. So in a way what we are doing is what I consider to be the last mile logistics and you can have the best product out there you may have a willing consumer for it you have an unmet need for it you order it but there is no delivery person so all of that goes down the drain. So in a way we are a little bit ahead of the curve because you know most people when they you know when we show these results for them sometimes you know people come up and say that okay but so what you know there is no treatment but we can't wait for the treatment to come and then decide okay now what are we going to do. So we have to do our job we have to take ownership of the problem and we have to try to deliver so that other things when they materialize everything is ready to go. In this world that you're describing in this success world where we really are bringing together all these different technologies in order to finally defeat pancreatic cancer. How are we delivering the right drugs and right screenings to the right populations because we're back at the first question that we keep circling which is okay what we have the screening technology that's assumed we had a screening it's 2032 we have screening technology that can find with extraordinary accuracy whether or not that tiny tiny nanoscopic little sliver of your pancreas is cancer or not cancer but we need to make sure that the right person is in the CT scan how would we do that how would we what would be the best way to make sure that we are delivering these drugs to the right population delivering these services to the right population. Absolutely so that actually is a million dollar question because without that question everything else becomes a moot point so here are a few things first and foremost there is a cohort there is a group of risk factors that have already been validated which is called new onset diabetes as we discussed with a high and back score and that's exactly the cohort that we are trying to recreate in our prospective trial. The question is a step ahead of that how do you in real time find out the individuals who have those risk factors so here is what we have done that may apply in case that you know the institution has given us funding to invest in those automated tools that screen your EMR the you know medical records of all the individuals who are part of that medical record in real time so which means that if I go today in the afternoon get a blood test get a fasting blood glucose get a hemoglobin A1C which looks at your you know blood glucose for the last three months and the results become available by 4 p.m. today and you know the system that we have developed and are optimizing now it looks at that value it looks at my values in the past and it tells the physician who is responsible for my care that this individual has got new onset diabetes and here is this impact score so which means that we are not trying to rely upon the manual work because that manual work is not scalable it is not automated and it can make a lot of errors so we are developing those tools some of those tools are already being used into this clinical trial that we are running right now but there is a lot more that needs to happen we still need to find out a lot more risk factors that can work for all the individuals but yes there is research that is already going there we are using our bioinformatics tools to be able to do that so it is already on the right trajectory. Do you think pancreatic cancer is going to be like lymphoma in 10 years a cancer that medicine knows what to do with? Yeah absolutely I have no doubt that that is where we are headed the question is that of course we'll have pumps along the road we'll need a lot more resources to be deployed but that's why initiatives like what you are doing right now which is that making people aware about what can be done what is being done is exactly the kind of conversation we need to be having. Dr. in this final set of questions I want to ask you a little bit about the issue of AI and radiology writ large I mean very famously among people who study the effect that artificial intelligence could have on the labor force it was nine years ago I think when the scientist Jeffrey Hinton said that quote people should stop training as radiologists now end quote and since 2016 when he made that or 2016 or 17 when he made that observation male clinic has increased its number of radiologists by about 50 percent you've hired an additional 400 radiologists so clearly male clinic itself is not adhering to the wisdom of AI godfather Jeffrey Hinton when it comes to shutting down the number of radiologists you hire because the AI is so good. This is a question that I wanted to ask radiologists for a while if AI is so good at reading images why hasn't it replaced more radiologists yet? Well I will answer that question but let me tell you what is the right question to ask the right question to ask is that how do we deliver the best outcomes to our patients if that process requires AI to replace radiologists I'm all for it but unfortunately that's not what it is the way I look at it is that the AI is a tool and as with any tool and technology its utility and its ability to make a difference is dependent upon the team that is asking the questions to it you know one of the analogies I can give you is that think of it like a smoke detector you know you need a smoke detector to tell someone that there is a fire in the house but the smoke detector is not going to go and get rid of that fire you still need the fire fighter to come in there to look at what is wrong and what needs to be done so right now that paradigm is exactly the same the kind of AI that is being developed that AI is a signal that tells you that there is something wrong but a physician has to make that judgment whether or not it is truly wrong what needs to be done about it and so on do you think it's possible that we are currently speaking at a high watermark of radiology employment like is it possible that the proverbial smoke detector gets so good that you can simply make do with slightly fewer firefighters as it is in this case that the same talented radiologists that in 1985 could do this amount of work can now do three times the amount of work very successfully and highly accurately in which case maybe you don't need the same ratio of radiologists to patients or radiologists to other doctors like can you envision a future in which AI gets good enough to in fact replace some radiologists or do you think it will continue to simply be a tool and you know tools don't replace workers tools enable workers to become more productive so is this the kind of technology that you see expanding to do entire jobs or do you think it will always be bound to tasks within the job of radiologists right so you know as someone who has work in this domain of radiology for a good number of years as well as someone who has had the opportunity to develop AI from scratch from ground up you know I don't foresee a scenario where you could have a tool that would make decisions especially in a high stakes domain like radiology which is what determines the downstream care pathway you know if we get it wrong anything that goes from there on is not going to be right so for this high stakes game given the complexity that is involved given the judgment that is required I don't really foresee a scenario where a tool or a group of tools would replace that judgment but again as I said you know we at Mayo Clinic are interested in delivering the best outcomes for our patients and what we will continue to do is to invest deploy and find even these tools that can help us do the same. My last question for you is a sort of word to the wise about artificial intelligence and medicine I'm very interested in how AI is being used to discover drugs, how it's being used to accelerate maybe clinical trials, and certainly how it's being used in radiology for diseases like pancreatic cancer. And I always find that with tools that are this powerful, there's going to be a good way to use the technology and a bad way to use the technology. A good way to use technology obviously is something like improving the accuracy of discovering pancreatic cancer by a factor of three, as this study suggested is possible. But I'm concerned that in a future where people believe that AI diagnostics are really powerful and accurate, that we enter into a paradigm where we are over-testing everything. Where lots of people are just getting full body MRIs and blood tests. And as you said, even if the sensitivity or specificity is 99%, if you're looking at at a population wide level, that still tens of thousands, hundreds of thousands, even millions of people who every year believe themselves to be at risk of or dying from diseases that they just don't have. And so that to me is sort of the Frankenstein worry of diagnostics gone wild. That's one concern that I have about artificial intelligence and diagnostics. But I'm interested how you see the balance of AI as good and AI as a warning in your space. Yeah. So you know, we are interested in data driven answers, right? So here's what we are doing. We are doing a study whereby we are looking at the dynamics when radiologists interact with AI on the scans that we have used in this study. And by the way, since the time of the study, we have increased that pre-diagnostic scans to about 550. And what we are doing is that we are doing studies whereby radiologists interact with those scans with and without AI to be able to answer those questions. But the year's a short answer, you know, AI is just a reflection of the human mind. The way we use it is going to be a reflection of human emotion, human greed, you know, human fear. So that is the reason why you need institutions and physicians at places like the Mayo Clinic. Because you know, we have been interested with the public trust and the responsibility to do it in a responsible and deliberate way. So we'll continue to do our job, but you know, we cannot control the world. Dr. Gowenka, thank you very much. (gentle music)