Man vs Machine - Are computers and AI going to replace dermatologists?

[music] Welcome to Derms on Drugs. A video podcast brought to you by scholars and medicine, the best educational platform at Dermatology and provided at no cost for healthcare providers, as long as you've got one of those NPI numbers. So, Derms on Drugs is where cutting-edge derm meets, nervous comedy. I'm Matt Zeyer, and each week I'm joined by my residency buddies, Dr. Laura Fares and Dr. Tim Patton, to discuss debate and dissect the hottest topics in Dermatology. We use our 60 years of combined Dermatology experience to keep you on top of what's really happening in the Dermatology literature. It is everything you need to know to be on the cutting edge and it'll be the most fun you've ever had while actually learning something useful about Derm. New episodes drop every Friday on Spotify, Apple Podcasts, Scholars and Medicine, and all other major podcast platforms. And I do want to remind people this is a video podcast, and so we do take advantage of that and use some of the visuals and tables from the articles that we're talking about. So if you want a little bit more info, check out the video, don't just listen. Well, today we've got a phenomenally interesting episode. So we are going to go into Skin Cancer Diagnosis, Assisted by Computers, which is an incredibly hot topic, and one that is really exciting to talk about, especially because we've got one of the world experts as one of the Derms on Drugs. So let's go ahead and get into it. And we are going to start by throwing it over to Dr. Patton. Dr. Patton, what do you got? Okay, my first D-dive is from the September 2023 journal of European Journal of Cancer, first author, Sermonara. It is titled, it's a long title. Stay with me. Diagnostic Performance of Augmented Intelligence with 2D and 3D Total Body Photography and Convolutional Neural Networks and a high-risk population for melanoma under real-world conditions. A new era of Skin Cancer Screening question mark. Did you say 2023? 2023. So like two years ago? Well, yeah, but I wanted to do something that was Vectra. That was like real weight. Okay, fair, fair. You're right. It's two years ago. But it does involve Vectra, which that was kind of what you were talking about as well. So I just was looking for another Vectra article. Okay, got it. All right. So the short answer I think to that question is not quite yet. It was a prospective single-setter study performed at the University Hospital in Basel, Switzerland. It was a study that wanted to compare humans to AI and the diagnosis of melanoma and then to also see how well augmented intelligence worked, augmented intelligence being humans combined with AI analysis of patients, Nevi. So January to August 2021, patients recruited from a high-risk melanoma clinic. So these were high-risk melanoma patients. Patients underwent a skin exam by a human dermatologist. And then we're analyzed by photofinder. That's a 2D total body photography device and Vectra 360 WB, a three-dimensional total body photography device. Dermatoscopic images of all Nevi, three millimeters and greater were made. And then they underwent a final exam by human dermatologist now armed with this AI information. This was the augmented intelligence that was described. There were two other parts of this study. There was a neat discount, which had some interesting results. And there was a subanalysis to see if there was significant deviation from two separate analysis by the same machine. I don't know if we'll even get into that. So this part of the study, excisions were performed if the derm classified it as a lesion, a malignant lesion versus benign. So it was very dichotomous, dermatology exam, malignant benign. The AI score delusions above a certain score. So if it was above like basically halfway, if it was halfway above their scoring system, then it got excised. Or it was the derms assessment knowing the AI score. That was a little confusing because like if the AI score was elevated, it had to be excised. So which lesions were the derms saying, OK, I'm going to excise this based on the AI score. My guess is it was like borderline stuff. They don't really get into that, but I think that had to be it. So 75 excisions were performed based on these criteria. This was out of a total of 1,690 lesions total on 143 patients. The table 2A is kind of where all the numbers are, sensitivity, specificity, positive and negative predictive value. The machines compared to human assessment. And they went through different like what was the gold standard. But if you look at histopathology, which I mean that's the gold standard, derms had the best sensitivity, specificity, positive predictive value, negative predictive value. The positive predictive value, the machines was terrible. 15%, 28%, a lot of things it recommended for biopsy that didn't turn out to be melanoma. Sensitivity of the 3D machine was same as humans, 90%. So both humans and 3D missed one melanoma out of 10 that were there. How did it find out it was a melanoma if both the humans and the machines missed it? Didn't nobody buy it? The 2D machine picked it up. Oh, so the sensitivity of the 2D machine was the worst. It was 70%. They missed three of the, there were 10 melanomas. 3D photography, humans got nine of them. The 2D got seven. One of those sevens wasn't melanoma that the 3D and the humans didn't get. None of the melanomas that were missed were anything bad. It was like in situ T1A lesions. The interesting thing I thought was the machines convinced the humans, like specificity of the humans dropped after taking into account the AI score. Meaning that the machines convinced the humans to biopsy more things than they would have otherwise because they must have had like close to a bad score. So they're like, you know what, I'm going to go back and I'm going to say these ones should have been biopsyed. So, you know, I, like this is kind of, I've had an end of two with the vector patients. Have you guys seen any vector patients? I have not. Yes, because when I was in Pittsburgh, an organization across town had it. So I would see those patients not infrequently several. Yeah, and you know, this is what happens is they come in and they're like, I went and I went and had this vector and here's the here's the vector of like paperwork. And vector would pick out a few of these Nevi and they'd be like, hey, you should look at these. But it's not like you didn't do a skin exam on those people, right? It was, you know, the, the, the machines did not find as many things concerning as the human dermatologist did. And it turns out the human dermatologist has, has a little bit of a better judgment. The take away I got so it's not like you can't more things can concern. And then the human dermatologist did, right? Yeah, way more. But also, you know, so like for there was one table 12 lesions out of these that was 1600 or whatever were considered suspicious. One of the units said, well, only seven of those 12 did we find suspicious. The other one said, well, only eight of those 12 did we find suspicious. So it just seems like I would trust my own sort of like exam over the computer. So the people came to me with a vector and it wasn't like I only looked at those lesions. I said, we're going to do a total body skin exam and you look and then I looked at the sheets and I'm like, yeah, these are all fine. We don't have to take any of these off there. They have a benign dermatoscopic evaluation. The, the, so you know, are these helping us? None of the vector patients that I've seen have I been like, oh my gosh, that's so great. You had this vector look at your skin. I didn't find it to be helpful. Now it's end of two and I think there's certain instances where. I'm going to tell you what changed and what didn't and maybe that helps. But anyways, the number of Nevi was also a pretty interesting thing. So humans identified a mean Nevi count of 210. I was like the gold standard 3D counted 469. That was the mean Nevi. And more than twice like what are they counting and even worse, the 2D device counted 1,324 Mevi. That was the mean number of Mevi that they came up with. That was like working with the worst resident ever like is this an EVIS? No, is this an EVIS? No. This has to be an EVIS. That's not an EVIS. In my own, my other favorite part of the article was the breakdown of the dermatologist into beginners, intermediate and experts. So who do you think saw the most patients in a skin cancer study? If your answer was the experts, you were wrong. Experts saw eight of the patients, beginners saw 103. I think that always cracks me up like I'm a skin cancer expert. And then you'd be like people would be like, Oh, you must do all the skin cancer screenings. And I think skin cancer, which I like. How though are you kidding me? So we're experts. You are the more you delegate. That's right. That's your test. That's why you're the experts. But that was, yeah, that was my paper. So your main takeaway from this was didn't really do not ready for prime time, was your takeaway from this? That's been my experience with the few patients that I've had. I was curious as to what this study showed. And it's kind of the same thing. You're just like, you're doing the total body skin exams and we just compared, like positive predictive value. We're just way better. They, they, they, they suspect a lot of things are being possibly bad that they're just not. Yeah. I, you know, I saw that paper in what I thought was interesting is if you look at that graph, right, and the lower left quadrant are the benigns and both systems, like the vast majority of lesions, both systems called benign, right? So I think it actually maybe AI is not, and I'll kind of talk about this. Like instead of looking at AI is can it find melanoma? Maybe we should look at it as like, can it just take off my plate of things to even look at the benign stuff? So it actually was pretty good. It like, it, like if you had just said, I don't even have to look at the things that called benign, you wouldn't have missed much at least with the vector. You wouldn't have missed about as much as the dermatologist did. Yeah. So, yeah, that's some negative predictive value. Really, really good. But I think that's also a function of the fact that the vast majority of things that we see are benign. Right. It goes down to like what, what's the most, what don't you want to happen? You don't want to miss the melanoma. So that's the sensitivity. And even, you know, you can have a bad, what do I want to say? You can have a bad false negative with a really, really high negative predictive value, because it's a rare event that you were looking at. It just takes one. Yeah. Yeah. So. All right. Let's, let's move on to the second paper because it, it, this one actually kind of blew my mind. So, this was just recently published in JAMA. So 3D Total Body Photography in Patients in High Risk for Melanoma, Randomized Clinical Trial. This was by Sawyer at Al and this was just in JAMA Derm published online March 26th 2025. And the initial takeaway from this in this study. So what they were trying to do, it was powered to detect, could it reduce the number of lesions that got excised? That are really just like define what excise means in this sense. I assume that they're including like a deep scoop that totally removes the lesion as an excision as well, but they powered it in order to be able to detect, hey, if the, if the machine can decrease excisions by 50%, that'll be great. So they had 314 participants. The, bunch of people in there and they were randomized 50, 50. So half got basically normal care. The other half got the 3D vector imaging with digital Dermoscopy every six months. So they came in and got re-imaged and got, you know, the lesions Dermoscopy as well. And basically what happened was that there were significantly more excisions in the people who were getting the 3D photography every six months combined with Dermoscopy and also fewer melanomas. And so my, my first take of this thing was, well, what the heck, it's just, they randomly got, you know, more people who didn't have melanomas, but they got in the intervention group, but they were doing more excision. So just telling people to cut off stuff that didn't need to be cut off. However, the interesting take here and the study was, so it certainly didn't find that it decreased the number of excisions. It increased the number of excisions. But the most likely explanation of this data and like the authors say, we like tired to know maybe it just randomly that were fewer melanomas in the one group than in the other. Now, these were all very high risk people who'd had, you know, at least had one melanoma before I think the age of 40 or two melanomas before the age of 65 or had, you know, a strong family history or had been diagnosed with plastic nevas syndrome. So super high risk people. And there were not an insignificant difference in the number of excisions, right? So the lesions excised per person in the, in this two year study and the people who were getting the imaging was almost six. So 5.73 compared to the control group, 3.99. And there were fewer melanomas detected in the intervention group. So 24 in the intervention group versus 43 in the control group who just got normal care. So at first glance, I was like, oh my god, this thing sucks. It just tells you nothing. It's not off. But a reasonable interpretation of this is that this whenever you combine 3D photography with their mosque, be with a dermatologist every six months, they actually caught pigmented lesions that were turning into melanomas. But before they would have been called a melanoma, before you'd call it a melanoma. So they literally, there's a reasonable possibility. And I back, you know, put it at a 50/50 chance that it was just randomly. There were fewer melanomas in the one group than the other versus the, whenever you do this really intense 3D imaging, digital microscopy with an expert reviewing the man to derm doing a full body skin exam every six months, you literally caught moles in ultra high risk people who were on their way to becoming a melanoma, but were not a melanoma yet. And that's why we saw fewer melanomas in the intervention group. Now, you know, that obviously, I mean, if that was the case, because it never in a million years occurred to me that you'd be able to do that. That there'd be like a way to catch moles that were on their way to becoming melanomas, but weren't yet in ultra high risk people. But man, if that plays out, that that's what this is doing, it really becomes a fascinating thing of like, who's got a high enough melanoma risk to warrant this kind of really intensive screening? You know, is it all of our melanoma patients from now on? Should be, you know, if it turned out to be true, we should be doing this, but just fascinating. And because it's, you know, we're going to kind of kick it over here to Dr. Ferris, our in-house world expert on all of this stuff, because it really may be think for the first time about what real, you know, I had kind of assumed that the goals of computer assisted imaging and diagnosis were to basically make it so that we only biopsy stuff that is cancer, so that if we get it to where it's got, you know, 100 percent sensitivity and 100 percent specificity because the AI has gotten so smart, that's the eventual goal. Never going to get to 100 percent, but as long as it gets to be as good or better than us, okay, that's the idea. But like the idea that it might actually get you to be able to cut stuff off before it's even a melanoma, and that's a hard thing to prove, but it becomes just a fascinating, you know, variability. So, you know, either of you have comments on this paper before kind of we kick it over to Dr. Ferris. Lots of thoughts on this paper. Actually wrote in a company editorial to the paper if you want like all of my vetted out thoughts, but a couple of things. One, this was a really weird study design. It was like during COVID, so basically patients had regular Durham was like you see your GP who was doing most of the skin cancer screening. And so you see your GP, they do your skin exam. That's the, that, and everybody has that. The experimental arm also had Vectra, but it was like some junior attendings, did the Vectra and then picked out the lesions that they thought was suspicious and then took dermatoscopic images. And then Peter Sawyer, who's like, you know, the world's expert on melanoma diagnosis, dermatoscopy looked at them and then advised whether or not to biopsy and then they sent that advice back to the GP who could follow it or not. So I'm just like trying to imagine this. You're like, I'm a non-derm GP and oh my god, Peter Sawyer is going to look at this patient. Now, because the patient knows what arm they're in because they know if they got Vectra or not. And so you're like, what do I do? I don't know about this spot. Like I don't want to miss a melanoma because Dr. Sawyer is going to be like that guy's kind of, you know, isn't going to think highly of me. I don't want to biopsy everything because Sawyer is going to be like, geez, that guy biopsies everything. I mean, I would hate to be that person. If you're the junior attending, you're like, do I? What's Sawyer going to think if I say this one suspicious? Right? So like, there's just so many, like, I was trying to think about like what to entitle my article and I was trying to put like six degrees of separation. I think that that didn't make the final cut. But it was like, there's so many degrees of separation from the real world use that I just don't know what to think. I am not thinking that it's. that we caught the earliest things. The other thing is this tool has AI, but in Australia, you're not allowed to use AI unless it's approved. So the AI features of the vector were not used. 'Cause they're not allowed to be. - Just think there's Australians who are so far behind. - Australia is Aussies. And then the other like fascinating thing is when you look at it, you're like the nine to malignant, the malignant excision rate wasn't different. And so the reason was you know what you found when you used Vectra, a ton of non-melonomous skin cancers. It actually really increased the diagnosis of those. So, and then there's like a cost accompanying, like cost analysis, and it definitely increased the cost to care. So, I don't know. Like there's so many variables at play here. I don't know like what the final answer is here. Pat, and what do you think? But did you see that article? What'd you think? - I did, and right, it was a big, I don't know. All right, I mean, so I, my personal experience with Vectra, it's like, I don't know, and my paper, it was like, I don't know, like, and then this one is, I have no idea how we're going to implement these, how good they are, how bad they are. It's really early, really early on in this process. We just have so much to learn. And the experience of everyone is gonna be so different at the beginning, I think. - How many centers use Vectra? Like it wasn't that, it's like 13 worldwide, or is that crazy? Just thinking-- - Maybe a little more than that. - But it's from-- - It's like 20, say, it's not tons and tons. - Yeah, I mean worldwide, it's just, nobody's using this yet. - Some very big well-known centers are. So, yeah, I mean, I think it's gonna be really interesting. And like they're mostly big well-known centers, but there are some community-based centers that do it too. So I think it's gonna be, I think it's gonna be very different, the use and performance of this, in the hands of community, non-derm run, sometimes centers, versus-- - Yes, and it's the expertise of the doctors, I think that's gonna be huge. It's also the patients they're seeing. You know, at some of these, the huge specialty clinics, it's gonna be these people that grow melanomas, once a month, versus, I mean, the people that are getting vectra, you know, what they have been diagnosed with melanoma, and they do have a fair number of moles. So that's the one I'd seen. - I mean, can you imagine this? Well, you know what, before I answer that question, let me get over to talking a little bit about Dr. Ferris's expertise in computer assisted diagnosis. So when the three of us were residents and very junior faculty, we were next door, essentially, to Carnegie Mellon University, which is one of the leading computer science places in the world, and I'm not sure how it happened, but Dr. Ferris ended up getting involved with them, and especially working with an early device called Melafind, which I don't remember how in the world Melafind worked, but we, you know, remember Ferris talking about it, and then one year at the AED, when we're all junior faculty, we go to some event where it's a chance for germs to like really see Melafind and how it works, and the whole thing, and whatever. We, you know, each have a mole. Oh, let's have a look at that one, and this one, and the thing says that patents got a melanoma on his back. - Well, that's right. It now, this was always the Melafind spiel. They're like, it does not say, if you have melanoma or not, it assigns a point system. And everyone, like if you went to talk, when one of those talks, you're like, well, yeah, and above a certain point system, you're gonna be like, I'm gonna biopsy that 'cause it could be melanoma. So they always had these silly arguments, but yeah, this one on my shoulder, that I had always had forever, and I knew looked terrible, but by the fact that I had had it for 20 years, I'm like, it's gonna call this thing bad, because clinically it looks bad, but we know it's not bad, because, so yeah, I wound up having to get a biopsy, 'cause I think Ferris was like, I work with Melafind, this thing finds Melanoma's, you gotta take everything off. (laughing) Way to go Ferris, good, driving unnecessary work. So now I'm physically and emotionally scarred by having worked with Dr. Ferris. And now we know why path is so-- But you're still alive, you're still alive. It was gonna hurt. You were evolving. You know, it could have turned into a melanoma if we hadn't left it on, if it had been left on exactly right. Ferris is your life. No, I will say, the downfall of Melafind was, I mean, I think multiple things, I think it was a bold thing to try to commercialize something to do this. But if you really look at in the end, what did Melafind do? 90% of the time it told you you should biopsy. - Yeah. - Of anything it saw 90% of the time, it was like biopsy, consider biopsy or whatever the high risk thing. So that's really not that helpful. If 90% of the time it's gonna tell you that you should get a biopsy, high negative predictive value, but very low specificity or positive predictive value. So in the end that was the problem. It also like had to go through, you like the whole classifier was fixed in the machine. So it wasn't like, it was a big, you know, expensive piece of equipment that had to get calibrated. It couldn't learn on the fly, learn from, you know, experience over time. You'd have to re, you know, have it like sort of re-approved through the FDA. So that was ultimately like the failure. So, well, my deep dives, I actually picked a couple papers to go over. So all, and then we can chat, but like, my first one is, I'm sorry Matt, also from 2023, but it makes an important point. Thomas said all, this is frontiers and medicine, real world post-appointment performance of a novel machine learning, based digital health technology for skin lesion assessment and suggestions for post marketing surveillance. Okay. So what is this thing? This is this device that's called Derm, which stands for deep ensemble recognition of malignancy. So what this does is this is like a thing, a dermatoscope attached to an iPhone that uses AI, and it classifies the lesion. So first, it's like, is it melanoma? Yes, oh biopsy. No, how about does it look like squamous cell? Yes, biopsy. No, how about basal? How about squamous side to AK, atypical nevis benign? And basically what it says is like, oh, this should be, you know, referred to, to a dermatologist, yes or no. So it's meant to be used. This is approved. It's approved by the UK medical device. It's a class 2A by the UK's sort of equivalent of the FDA. So this is out there. This is being used in the UK. And they'd say the national health system is so overwhelmed with referrals to dermatology for suspicious lesions. We cannot see everybody. So the idea is like, can we use this as a way to triage? So they have developed this. We actually did clinical trials with this in Pittsburgh. And now it's sort of being deployed there. So what happens in what this sort of first real world use was you've got a teledermatology hub. You've got a suspicious lesion. You go to the teledermatology hub, the GP, takes a photograph. You get the machine says suspicious for melanoma or malignancy, yes or no. If it's high risk, then it gets reviewed remote. So it says high risk, low risk. Everything actually then goes and gets reviewed by telemed by derm. And then if the derm says yes, it's suspicious, then it gets, then that patient goes in for evaluation. It can be biopsyed, yes or no. And this was sort of like failsafe to launch it. So it's reasonable to think of this as instead of the primary care doc referring to derm, it's like ordering an x-ray. And then if the thing comes back, you know, the machine says it looks a little funny, then a derm who's now acting as the equivalent of a radiologist says, yeah, we really do need to see that in person, but if the computer or the derm are like, nah, that thing's fine, then it doesn't. Right, then the patient doesn't ever come in. Does it, and it doesn't, do they pick, does the primary care doc pick like, I want this one, this one, and this one image, or does it do like a full body thing like that? It's not a full body. So this is like, I think one of the important things to think about with AI is do we want AI for full body or do we want AI just for lesions that are of concern to the patient or the doctor, right? Those are two different things, and they both have pros and cons. And so derm, if you look at like their version B that was deployed at two different hospitals, basically, you know, which is there, I'm just going to talk about that because that's sort of their most optimized one. So they say they have a sensitivity for melanoma of 100%, and a sensitivity for skin cancer are not of like, depending on the version, 99 to 100%. Specificity for melanoma is pretty good. 80% is about 60% for skin cancer, yes or no. Negative predictive value is 100 to really like 99.9%. I actually saw their most updated information at a late breaker at AAD. So they say like our negative predictive value is 99.9. For dermatologists, it's like maybe 95%. So they're saying we actually have a higher negative predictive value. Positive predictive value is lower. It's like 12%, 17%. in their study. So what I thought was interesting was they sort of went through this whole algorithm of you know when they deployed it at these two sites and so you know the UH the first site had 4800 cases where it was looked used in 3600 and not used in 1200 not randomized. It's just like how it shook out with how busy or whenever they were that day. The other site had 1400 cases roughly 1170 of them were assessed and the other like 240 were not okay. So you sort of have a non randomized comparison group and so what happened with those two groups they 41% at one site and 25% of them at the other site. They were the machine was like this is fine. This is benign. So then they got referred you know by a second reader and you know some of those then got brought in to be looked at basically even though like some of them were evaluated the Durham said look at it 0% of those were skin cancers. So 100% negative predictive value. So you know the message there was like you could have actually trusted it and not seen it of the ones that were sent for review. It was like 58% in one group 75% of the other. First of all very few of these were actually like on telemedicine did they say no don't worry about it. Most of them they did say come on in and see them of those that were biopsyde like 15 and a half and 17% were skin cancer and 4 and 5. I'm sorry 15.5% or 17% were actually skin cancer so that was like the positive predictive value and then 5.5% of those or 4.1% were melanoma. If you look at the ones where Durham wasn't used they just like went to telemed and then most of those they said come on in we will take a look at them the positive like that what percentage of those were skin cancer in one group 1.8% were skin cancer. There was two melanomas which was 0.17% versus 0% in the other groups so what does that mean you know potentially if they had just and now actually it is approved for use this way I just read like this could just weed out 25 to 40% of the of the lesions and say that doesn't need to go it doesn't need to see a Durham period doesn't need to go to a doctor. And then it could sort of route then if you just said we'll just see all of those ones that were actually read as positive you know your positive predictive value might be would be a little bit better so you'd sort of have. I think it takes the workload off the Durham but it still keeps in it's like what can I safely say doesn't need to be seen but then you don't buy up see everything you just say then use the Durham expertise and have them say yes that needs a biopsy or no so I thought that was kind of interesting. How does it actually work in Britain then these Durham machines are like set up at the subway there they're in a doctor's office with the GP this is not like in a in a subway. Is it in imaging is it in like the radiology department of a hospital or is it in an actual like I think it's like in an actual like GP clinic okay. Yeah like they get lots of you know GPs are doing more more skin cancer there but you know I thought kind of an interesting idea that you could just triage and say that doesn't need to be seen. The other cool paper that I wanted to talk about was just published online and jamma Durham on April 9th and this was this was. This is a Spanish group so Rose you keep air which I'm sure I'm saying completely wrong this is this is. Standard dermatoscope images versus an autonomous total body photography and. Dermoscopic imaging device okay so this is like the next next generation vector almost next generation vector so if you look at I mean like if you're interested in this stuff you totally have to look at this paper so one if you look at what the device looks like it's not this like look at the vector device it's a it's you know 90 something cameras all over it's very cool you stand here this one. Is a little bit more like what looks like a big butterfly wing device and it goes up and down and scans the patient what's crazy is it it takes you know clinical images but it also takes non contact. Dermatoscopic or as they say dermoscopic images of every mole so if you look at the pictures you're like oh that's dermoscopy that was not like a person with a handheld device that was the machine. So from a distance from a distance it can you wow yeah so like you have to go look at this paper because you'll be like here's two images which ones the better picture and you're like I think that one's better and you're like oh my gosh that was the one that the machine took it actually to me in a couple examples they showed like a better dramatic or a better dramatic image but then what they did was they basically did they they had a reader study and they said is this they took each one and they and they didn't they blinded them to if it was like taken by the machine or by a person with a dramatic scope and they said is this good or bad and basically what they showed was it was they showed non inferiority of the machine and but if you look at each one the machine actually had a higher percentage of lesions where they said this is a good image it was 98.44 to 100% depending on an atomic area as if you looked at that for like the standard dramatic scopic images it was like 92.55 up to 96.42 so actually like readers thought that the images look better now there was zero melanoma identifications can cancer detection in this paper but if you can imagine you could scam somebody head to toe you've got there you've got the clinical image you've got the dramatic scopic image and then you can really train AI on that like I think this is super cool it's not like pick which one to image it's not you know Peter Sawyer's junior attending running around taking the images and then he goes and it's like you would just have all that done automatically so so this really could replace these terms so potentially if you could really image I think it could like this could be the remote total body skin exam now let's say you combine this with AI and you say what you're going to do is get imaged every six months and then the computer AI is going to be able to say here are all the lesions that changed like by based on you know clinical images here's all the ones that have changed dramatically and then you know you have a human look at the ones where there's enough change to cross the threshold like to me that is where it could be interesting because I think the number one best predictor of biologically relevant skin cancer is change if it's not changing it's not biologically relevant skin cancer if patents mall on his shoulder had not changed in 10 years even if that pathology said 0.1 millimeter melanoma that was not a lethal lesion it was a histol lesion with histologic features of melanoma but if it's not if there is no cell division there's no change I think that's not a Malig I think that's not a potentially lethal biologically relevant and if this thing is taking high quality magnified dermiscopd or metoscopic images we're really going to be able to see even minute change so you'd really be able to say that thing is not changed really at all in the last you know six months five years whatever wow wow so all right Ferris let's let's I want to so you you are one of if not the smartest person I know and you've been thinking about this and working on this for 20 years now and what the like what do you think that the realistic goal is is it that we'll get to where you know in the next couple of years devices like this make it so that people can get remote skin cancer assessments and so that as a derm everybody who walks in you're doing surgery on because they've been you know whatever AI imaged is that what the goal is and if you know if it is you also or somebody who now runs a large you know the equivalent of a large group practice right as a or like I like to call it an academic department but yeah yes but for for you can think of it but you still have to meet your budget numbers right and if like is this the terms need to be worried because this is going to get good enough it's going to get good enough that we don't you they'll be able to send them to get image they want to send it to see us is this going to like tank dermatology I do not think so. So, you know, I started in this thinking, what is our unmet need? Oh my gosh, we need to be better at finding early melanoma and we're gonna save lives. And that's the most important thing. I feel like we have, you know, our workforce and access issues are very serious problems in most of the country. So, you know, and I think why do we have such access problems? I think our access problems are very much due to skin cancer screening and not skin cancer screening in high-risk patients, but sort of the general population interest in skin cancer screening. Well, you know, skin checks, regular skin checks for people who are not at high risk of melanoma. And so, I think that that absolutely can take up most dermatologists day. I think that that means that the people who get screened are the people who are in generally very health conscious. They're probably also the people who should they develop a melanoma are more likely to see it. It's probably not older, rural, you know, poor males who are the people who are most likely to die from melanoma. So, I, to me, like, I look at this as a way of saying, it's like to say to breast oncologists like, oh my gosh, we have mammography, like the breast exam isn't gonna be relevant anymore. Are you worried about your, you know, future? I'm like, no, like you don't want to do breast exams on people who don't have, you know, breast cancer. You want to actually find the people who have breast cancer. Ideally, you want to actually find the people who have biologically relevant breast cancer and you want to treat them. And I think that there is so much for us to do. We have so many diseases to treat. There's still so much skin cancer out here. I'm not worried about that. My hope is that this will actually let us focus our attention on the people with lesions of concern and at the very least, imagine that a patient comes in and they get scanned and you can stay, don't look at every single mole on their body. Here are the four to focus on. And then, you know, combine that with our expertise. So, you're gonna find subtle dermatoscopic changes, subtle new lesions. I think our expertise is still gonna be incredibly important in making the right decision. And we need to not look at technology as, oh gosh, it found it, I better biopsy it. I think we use it as, you know, instead of identifying that one malignancy, can you just find those 99% of benign lesions that I don't even have to spend the time looking at and pulling out my dermatoscope and, you know, looking like I'm doing this totally thorough exam, imaging everything, even though, you know, maybe I know that's not what it is. Pat, what do you think? What are your thoughts? - Do you think that there are practices that say, yeah, you know what? Like, you know what is a huge money revenue generating process for us is skin exams on 100 people a day that actually don't need it and we're not gonna ever find anything, but man, we make a ton of money off of that. And you're right, even if you don't biopsy them, right? If it's a five minute visit and you get a 99213 that's worth 120 bucks, if you do a lot of those a day, you know what, I am optimistic about our colleagues and I actually don't think that. - I don't think that either. Like, if there are practices like that, then yeah, they'll be decimated by AI. - But do you, I don't think there's any practices that think that, that are like if you hook them up to a lie detector test. - Say that again, you cut out first time. - So I don't think there are any practices that think that way, that if like if you hooked everybody in every German America up to a lie detector test and said, do you use this, they don't be like, no, no, no, no, no, no, I don't know. But they don't realize that that's what they're, you know, the, you know, a large part of their practice is. - Right. - They wouldn't say we, I had 120 skin cancer screenings. I averaged that many a day. They would say, look at what great care and how - I were providing access and reassuring. - Yeah, and if you hook them up to a lie detector, yes, they would probably completely pass. - So I, yeah, so I guess what I think one Matt, thinking back to your paper that you talked about, what are we gonna find? And this is probably saying we need to think about, we're gonna find tons of little basal cells with this, right? Like, you know, the longer you're in practice, you're like, oh, I can like see a three millimeter basal cell a mile away. They probably didn't, I didn't need to find that on an 85 year old, but now I've got it. What do I do? So if somebody's really worried about this, I would guess that their non-millanum skin cancer excision rate is gonna go way, way, way up. So that's one thing. - Do you have any sense of how many people in America get screened on an annual, get a skin check on an annual basis? Is it 30% of the population? I've never seen this number anywhere. I don't know if it's like a super secret number that nobody knows, but you, you special people, like, - I could tell you, but I'd have to kill you. - So, no. No, I, so we do not, I'm actually like one of my great academic classes I would actually like to know that number. The problem is, as you know, we don't really like code for it. And so we don't, like, there isn't a code for skin cancer screening. There's like a Z or V code that you can use, but not everybody uses it. So I've tried to look at this. There are sort of validated, like, ways to look at claims data to say if there's, you know, a 9.9, and this was more before, like, we changed coding, but like a certain visit level was certain diagnoses, like sub-retaritosis, AK. You could say that that's likely a skin cancer screening examination. So I actually published a paper also in the year 2023 with McKenzie-Warner and in, who's at MD Anderson, using Medicare claims data. And what we found is we just said, what percentage of Medicare visits with dermatologists are likely for skin cancer screening? And in 2018, it was about 73% of derm visits in the Medicare population. In 2009, it was more like 53%. Okay, so we're doing more and more of it. So, you know, that is one thing. And then the other paper that I had was, there's something called the National Health Interview Study, which is where basically, like, if you're a doc and this is across all specialties, you know, they'll say, hey, would you participate? Like, we're gonna basically give you a couple days or, you know, sessions and say, tell us for each patient you saw, like, what was it for? You know, answer these questions. And so we looked at, like, you know, changes and rates of total body skin exam over time. And so looking at their data. And we found basically, it does seem to be increasing, particularly among women, younger people, wealthier people and people who are more educated. So we are doing more and more skin cancer screening. I don't know, like, what I'd love to know is what percent of our workforce effort is dedicated to skin cancer screening. I don't know the number, but I think it would be really interesting to know. - And you have reassured me. So what I'm picturing now is, right, here in Ohio, if you live in Columbus, if you wanna get screened for skin cancer tomorrow, like, you can get screened for skin cancer tomorrow. There's no shortage of visits. But if you live an hour outside of Columbus and you don't wanna drive an hour, even if you're primary care doctor, like, you really better get checked. Like, people, there's a good chance they won't. And if we get to where there are devices like this, living in radiology departments or wherever. And, you know, Joe, 80-year-old can, like, go over to the local hospital and get a full body thing. And then it says, this is a skin cancer. You need to go get it treated. - Or this is just suspicious enough to warrant a visit. - Yeah, it'll make it easier to get people to follow up. So it may significantly increase the number of people getting skin cancer screens. And these devices should be very cost effective. I mean, you know, we pay for at once. And then it's, you know, it's not like you need, it's not an MRI machine or a CT machine. You know, it's a fancy camera. - Yeah, I mean, there's probably, everything is always more expensive than we think it should be or, you know, than we would imagine. So I'm sure that there is that. There's gonna be like, you're gonna need a person to coordinate it. But, you know, I think the big things are like, one, can we save the time that we spend, like, doing a manual physical exam head to toe looking for stuff? Like, that's just, it's a, you know, every year on every person is not feasible. It's probably not the best way to do it. So could we then say, here's what the machine picked up? Maybe then with good images and sequential. And so detecting change is key. Let it dermatologist look at that. Then if it's suspicious, then it goes in and gets evaluated in person and can get a biopsy. I think like, you know, you would imagine maybe we would actually then be able to evaluate twice as many people. Hopefully we would limit it to the people who are actually at high risk. And it's not everybody in the world. It's truly the people at high risk of, you know, getting or dying from that. melanoma. So putting change in as a feature and cutting out the step of a human being looking at every physical square millimeter of the skin, to me that makes a lot of sense. And then in terms of like the cost thing, like, you know, I do think a lot about this and I, or like what is our motivation? I really think the problem is whenever you do a, you know, when you buy up CML, melanoma on a patient, those are your most grateful patients, right? They really feel like you've saved their lives. It's hard to them be like, nope, that we probably didn't need to do that, you know? And then it's also hard to say you don't need to come back. It's also hard to say, I don't need to see your neighbor for skin cancer screening. It's also like we've all lost patients to melanoma and it's hard to say, oh, we don't really need to go looking for that, right? So it is reassuring to us when we find it. It's very, reassuring to patients when we find it. Patients are really grateful that we did and do a biopsy. They're often grateful that we did do a biopsy. I think it's, I think it's a really tough problem that is not driven by greed. It is driven by like our desire to do no harm into, you know, why would I rather do a biopsy or lose a patient to melanoma? I totally get why we're in the predicament we're in. Yeah. Patent any closing thoughts? Nope. I agree. I don't think it is a great thing. I think it is like, you know, when you talk to people that do a lot of biopsies, they'll, you know, tell you the story. Well, I didn't think this one was and it turned out to be and I don't ever want to be in the situation where I didn't biopsy something and it turned out bad. So I think it is driven by concern for patients not like, ah, this is a huge money maker and I'm going to biopsy a lot. You know, it's something I learned whenever I left academics and went into private practice and, you know, I'd always thought boy, these people see so many patients because they're just, you know, they mean the more people you see, the more money. And no, it's that patients like there's a need and, you know, saying to somebody, well, I want to have 15 minute or 20 minute appointments, but you're going to have a 14 month wait to get into see me is just not good care. And so it, yeah, the it okay. I agree with you guys. I don't think our colleagues are greedy bastards. All right. In patient care, the people that go out and do speaking and ad boards, greedy bastards. I don't know if you two do that or not. I have no idea. I don't want to. I know you don't do any of that either. We're good. All right. Well, that has been a fantastic discussion today. And now let's throw it over to everybody's favorite segment of the show. The patented patent trivia. It's AI AI themed. I like my chances this week. I pay a lot of attention to AI. All right. Deep blue IBM's AI system defeated Gary Kasperov and Chess in 1997. But it took artificial intelligence systems until 2016 to defeat the world's top player in this other strategy board game. Go. Yeah. It's called go. It was Lee Sedol was defeated by Google's Alpha Go system. That was the big thing when deep blue beat IBM. They're like, well, there's this other game and AI is never going to be good enough because it takes thinking and strut and creative. And AI like AI can handicap itself now and still beat the world's best go players. It's kind of crazy. All right. Uh, number two, this onboard AI system was not at all helpful to the crew of the spaceship in 2001, a space Odyssey. How? No. Yes. Technically how 9,000, but we'll go. I beat your fares. I got there first. He said he said. Uh, funny thing when I was reading about this one, people thought how HAL the reason they came up with that name, it's because it's one letter before IBM. That's not true. That's not true. It stood for heuristically. So H programmed algorithmic a L computer. That's where the name how came from. I like the idea. I like that story too. It's just better. Yeah. Yeah. Yeah. Number three. What was the name of the IBM computer system that defeated Ken Jennings and Brad rudder in the televised quiz show jeopardy? Deep blue. No. No. Oh, it is Watson. Sorry. I'm not a lot of second guests. Evan. No, you're allowed to second guest. Okay. You're good. The only one you couldn't second guest was the, the, uh, bio, bio similar name one. Okay. Okay. Okay. Otherwise second getsons. My first man and that's on a role. You are. I know. My win. You know what? I was going to do AI in movies, but like Matt watches movies and Ferris does not. So I'm like, okay, I'm not going to do that. I do want to hand cap Ferris, but man, that's two weeks in a row. Zayers has come out. I think last week I think it was just a tie. I don't think I won. I think you won. Wow. Okay. I'm getting smarter. Hmm. I could have ever guessed that could have happened. Well, everybody, I want to thank you for joining us today and getting to see our world expertise in action with Dr. Laura Ferris teaching us that we don't have to be terrified of the computers. If you've got questions, comments, thoughts, ideas for topics you want us to see on the show, shoot us an email at [email protected]. Again, that's [email protected]. And we'll hope you learned a few things. Hope to laugh once or twice and mostly we're hoping you're planning to join us again next week. And until then, I'm Matt Zayers. I'm Tim Patton. And I'm Laura Ferris and we are Dermzondrugs.