Tech Check: AI and Emerging Technology in Dermatology Practice

Talk of artificial intelligence (AI) seems inescapable, but are AI and other emerging technologies really affecting daily dermatology practice? David Cotter, MD, PhD gives a status check and preview of things to come. Plus, Hilary Baldwin, MD shares tips for sunscreen recommendations and David Cohen, MD addresses new data on biologics and pediatric growth rates. 

Like what you're hearing? Want to learn more about the Dermatology Education Foundation? Explore assets and resources on our website.

Transcript

This transcript is provided as a courtesy and has not been edited for accuracy or clarity.

Welcome to The Dermalorian Podcast from the Dermatology Education Foundation. It seems like artificial intelligence or AI has exploded into the popular consciousness, with implications for everything from online searching and shopping to advanced aerospace research. But are AI and other emerging technologies affecting daily dermatology practice? To find out, we turn to dermatologist, Dr. David Cotter. Dr. Cotter practices at Las Vegas Dermatology and is on faculty for the DEF Essential Resource Meeting or DERM. He gives a preview of things to come.

Dr. David Cotter:

I hope you believe me, that the future is now. It's here. Things that we used to think of as science fiction are now available in our practice. We've come so far from De Morbis Cutaneis, which just had its 301st year anniversary, written originally in 1723, this treatise of diseases of the skin. To the development of the microscope in 1600, to the 1980s when we had the introduction of dermoscopy. It's quite commonplace, but rewind, even just 20, let alone 30 years ago, we weren't.

In the early 2000s, we see reflective confocal microscopy, RCM being utilized in clinical practice to diagnose melanoma and non-melanoma skin cancer in vivo. They're cumbersome and they're expensive and they take a long time. But we've moved beyond these fantastic technologies with new imaging modalities that can be coupled to machine learning to eventually give us bedside diagnostics, scan an image, get a malignancy prediction score. That, along with molecular technologies really are the future of medicine, and we're going to get into that today.

But first, we'll visit the realm of advanced imaging in dermatology. You may be a little familiar with RCM, maybe you've even heard of optical coherence tomography, which allows us to visualize lesions in vivo. That was a new technology when I was a PhD student about 15 years ago, but now you can actually use it in clinical practice. There's a more advanced version of that called dynamic OCT, which allows for visualization of blood flow in vivo. That's called optical coherence angiography. And the exciting thing about these modalities is they can be combined with one another. You can combine OCA with multi-photon microscopy, which is a type of fluorescent imaging, which allows you to look almost at an H&E style section in vivo. Millimeter wave imaging is the one I'm most excited about, and we'll talk about that shortly.

But to give you an example of how these technologies can be leveraged in our practice, a nice publication that looked at in vivo multimodal imaging of dermoscopically equivalent melanocytic skin lesions. So you pull out your dermatoscope for that funny brown spot, you identify 10 banal nevi and 32 melanomas. Great job. Your diagnostic accuracy is wonderful. Out of the 60 that were your difficult to diagnose melanocytic lesions, you got 42 off your dermatoscope alone. But what about those 18 that still remain difficult to diagnose? Well, in this study, six ended up being banal nevi, 12 ended up being melanomas.

So what the investigators set out to do was use multi-photon microscopy and OCA, optical coherence angiography, on the known images, the 10 nevi and the 32 melanomas to generate some data, some visual image data from which they can use principal component analysis to extract out the most discriminating features between the nevi and the melanoma. And they come back and they map their unknowns against those algorithms. Polymorphic cells, if those are present, the odds ratio of that lesion being a melanoma is 163 times more likely than it not being melanoma. That's what a dermatopathologist is looking at. That's part of the gestalt when they render a diagnosis. But this is being done in vivo. We're finding these structures with MPM imaging.

Taking it a step further, we can look at features that don't discriminate. Features like normalness, big deal, they're present in melanoma, they're present in nevi. Who cares? So the computer can take that information and use that as a negative feature that's not carried forward in their algorithm. So both positive and negative features are helpful when it comes to machine learning. If you didn't know it already, hear it here first. Melanomas are vascular lesions. It's not just your basal cells that have lots of blood vessels. It's just that the pigment in some of those melanomas make it so they're hard to see. But next time you have an amelanotic melanoma in your clinic, take a good look with your dermoscopy and don't push hard because you don't want to blanch it out. You'll see a milky, red homogenous area that corresponds to all of these blood vessels. And melanomas, in fact, have a higher density of small blood vessels, a higher density of large vessels, and an overall greater vascularity.

So you take all of your known nevi, all of your known melanoma, and you map these discriminant functions to find your secret sauce. And in this case, the secret sauce is a plot of discriminant function one, discriminant function two from this study that results in clustering of the unknowns. So you take your unknowns and plug them back into the analysis and you plot them based on the discriminant functions. And we see that our benign lesions cluster very well, as well as our melanoma and situs and our invasive melanomas. And in fact, the percent correct, depending on the thresholds you set, is 100%. Pretty exciting. Not quite ready for prime time because we obviously don't have these technologies in our clinic, but stay tuned.

Speaker 1:

Definitely stay tuned while we pause for this episode's Dermalorian Clinical Clip. It's Skin Cancer Awareness Month, and the challenge of persuading patients to use SPF continues. This can be particularly true for those with rosacea. Dermatologist Dr. Hilary Baldwin spoke with us at a recent conference to share her best advice for encouraging patients with rosacea to use sunscreen on a regular basis.

Dr. Hilary Baldwin:

So my best advice for rosacea patients who have sensitive skin and also need a sunscreen is pick a sunscreen that doesn't bother your skin. I know that's a ridiculous thing to say, but the bottom line is that you don't know what's going to bother the individual patient. So there is a little bit of trial and error involved. So my best suggestion for us as derm providers is to go into your closet and gather up every single sunscreen that you have, put them all together and let the patient give it a try so they don't spend a million dollars looking at sunscreens. In general, the physical sunblocks seem to be less irritating than the chemical ones, but that's not always the case. So unfortunately, trial and error is where I am in my practice.

Speaker 1:

Those samples can also come in handy for patients who travel. Speaking of which, let's rejoin Dr. Cotter as he discusses the diagnostic role for millimeter wave imaging.

Dr. David Cotter:

That's the screening that they use at TSA to make sure you're not packing heat as you walk through security. So what it is here though, they've set up these radiating antennas that scan the lesion of interest on the patient, and that generates a signal that then goes to a computer that generates these three-dimensional millimeter wave images. To a computer, they're quite meaningful because what the computer is able to do is it takes these three-dimensional images, which have data points in X, Y, and Z dimensions, ultimately generating approximately 1,000 different variables. And it filters them out and distills them down to the principal components using a variant of discriminant function analysis. And they take over 1,000 initial variables and they filter it to six different principal components they feed into the classifier model. We haven't really talked about AI yet, but that's what's happening here.

But what this classifier model is, there are five different ways to analyze the principal components that were derived from the 1,000 different spatial variables from the images that are generated by your TSA screening device. And they pop out a malignancy probability score, which gives you a binary endpoint of precancerous or cancerous tissue or healthy tissue, read biopsy or not biopsy. Millimeter wave imaging devices can be powered with a battery that would fit inside your iPhone. So imagine you're in clinic, instead of pulling out your dermatoscope, you pull out your millimeter wave imaging device and you scan the lesion on the patient. And within that device is an algorithm that gives you a malignancy probability score. So instead of in your mind having your dermoscopy gestalt, or maybe even a dermoscopy algorithm, presence of this or that, you just scan it. Talk about simplifying the process, especially for difficult to diagnose lesions.

So how well does this perform? It performs extremely well. This is a receiver operating characteristic curve, and the area under the curve tells us how well the test performs. Closer to one means closer to perfect. And certainly we do see separations between BCCs and SCCs and melanomas from other benign lesions. So stay tuned. Maybe you know, one year at DF, instead of giving out dermatoscopes for your attendance, we'll give out millimeter wave imaging devices. It's coming, guys, whether we're ready for it or not.

But what about artificial intelligence in dermatology? Everyone has heard about it with ChatGPT and other AI platforms, it's infiltrating into all of life, not just dermatology, not just clinical dermatology, but even in your reimbursements. So if you haven't heard, insurance companies are scanning your notes with AI and deciding to pay you less. AI can be used for good and it can be used for evil. But when this paper dropped in 2017, Deep learning algorithm does as well as dermatologists in identifying skin cancer, you'd think an atomic bomb had gone off. Shook waves. People were scared. They were, I think, insulted. "How can a computer do just as good as me? I spent all these years training. There's no way." Well, let's unpack that. Let's see.

Is it true the Industrial Revolution is over? The machines have won? I'd argue not yet. Not yet. Because the machines still can't tell the difference between the marker and the lesion. Which one are they supposed to analyze? Which one is supposed to be identified and input into their model? They're more worried about the gel, the ruler mark, ink marks, water bubbles. These types of artifacts are things that the human brain automatically disregards, automatically disregards. Machines have to be trained to do that. And not only do they have to be trained to do that, they have to be trained to do it in a way that doesn't exclude important positive variables because maybe you do an image, you take a dermoscopic photo, an angioma and a pigmented lesion. Which one should the machine identify?

So we have to proceed with caution here. I don't think skin cancer detection by AI is ready for prime time yet, but let's understand how it works so we can make an educated assessment when these technologies show up in our clinics, when the reps come in and say, "Hey, we've got an AI platform. We want you to use it. It's going to help you." Or better yet, when a patient comes in and they say, "I scanned a picture of the spot on my leg, they told me you should biopsy it." Or maybe, more fearfully, patient scans a melanoma, and the AI tells them not to biopsy it.

Well, it all starts with image acquisition, and then they have to process that image, and then they have to segment it, extract the features, and then classify it. Segmentation and classification are the most important processes in artificial intelligence identification of skin cancer. And it turns out that for those two deep learning models work the best. So how does deep learning work as it relates to identifying a skin cancer? Here's how we think about it. There's an original image, that gets uploaded to some type of platform. The computer then identifies the lesion, localizes its boundary, they crop it, they resize it, and they normalize it to generate an input image that can go into the AI model. Typically, it's a convolutional neural network that extracts some deep features. And those deep features are fed into what's called a classifier model. Think of it as a filter. You start with a whole bunch of apples. You start juicing them, you're squeezing them. There's some pulp, there's some skin, there's some seeds, there's stuff that you don't want, but you want that juice. So you filter it through a classifier model to get the good stuff out at the other end. And in this case, the good stuff is a binary result that says melanoma or not melanoma or biopsy or not biopsy.

This is what the computer does. It's a quite complicated iterative process, and the best model still doesn't exist. Imagine taking a single image and analyzing it thousands of times, multiple layers of data are extracted from it that feed in to the next phase of the analysis. To give you a real-world example, if you were in high school and you were doing an AI class these days, you might design a program to identify a dog from a picture of any animal. So you'll start with a picture of a dog, you feed it into the system and you filter it based on wet nose, computer learns a dog has wet nose, maybe the next filter is no wings. No wings is set as a negative. So dog now becomes wet, nose no wings. Wet nose, no wings, four legs, furry tail, likes people, so on and so forth. Can't breathe underwater.

You do that thousands of times generating multiple layers of data, multiple variables that then the computer can pool. So they distort the image to generate these multiple variables. They filter them to learn what's positive for the disease of interest, what's negative, and they pool them to make the analysis easier. So you lump together things that run together, such that eventually after thousands of iterations, maybe thousands of pictures of dogs, thousands of pictures of other animals, you can feed in a picture of a fuzzy four-legged friend with a wet nose wagging tail, absence of wings, can't breathe underwater, and it's going to tell you that's a dog. That's essentially what they're doing when they're trying to identify melanoma from a banal nevus.

But there's more than one way to skin a cat. So even though conceptually we might be able to think about it, people argue over what the best model is. What type of convolutional neural network should you do? What type of classifier model should be coupled to that? What type of computing power do you need? How long will it take? So on and so forth. To take a look at one real world study that gets at what would have clinical utility for me and probably for all of you would be imagine if you had a patient who had a broad field image. They took a picture of their leg and they said, "I want to know if I have skin cancer on my leg."

That's exactly what these investigators set out to do. They take that broad field image and they plug it into a lesion identification convolutional neural network that ultimately finds a lesion of interest that's cropped, resized, normalized, and now fed into their malignancy predictor screen. That sounds great. That's going to make our jobs a lot easier, right? Well, maybe not. Because even though red and blue seem to cluster malignant from benign, there's unacceptable amounts of overlap. These blue dots, the dark blues are basal cells and they track all the way throughout. There's even melanomas hiding over here along with banal nevi. So it's not ready for prime time.

And in fact, we've used some AI platforms in our practice to aid in skin cancer identification for people with multiple melanomas, multiple atypical nevi. There's many of these AI platforms out there, and sometimes they're more trouble than they're worth. Here's why. You ask the AI to find an outlier, it depends on what sensitivity threshold you set your outliers to how many outputs you get, and then you're trying to map all of these outliers that come out from total body imaging and find that same spot on the same patient and figure out if you think it's an outlier or not too. So now we have a medical-legal problem where computer says bad provider has to say good or bad. And the reality is you get so many outputs, you can't do 100 biopsies on a patient. That'd be unethical and unhealthy.

In addition, if you were trying to take a patient over time and you start with baseline imaging and you have follow-up imaging and you ask the AI to compare them, if you're off even by a few millimeters in the angle of your photo, depending on the sensitivity, you're going to get a lot of new outliers, a lot of new changing lesions that then you've got to go in manually with your dermatoscope and look at everything and ask, are they really actually changing? So what's the verdict? Not ready for prime time, but stay tuned. We need to lean into this because it's coming whether we want it or not. We can't just stick our heads in the sand like ostriches. Tech people are smart and they see an opportunity to make money in our space, and we need to understand these technologies so we can help our patients and be good stewards of their care.

And as we wrap up, I want to talk about what's most near and dear to my heart, molecular testing in dermatology, which literally is just the testing of molecules in the skin. We've got a few different molecules to choose from, DNA, which can be analyzed from anything from fluorescent in situ hybridization to CGH, SNP arrays, and next generation sequencing. What these do is they detect hardwired abnormalities in the DNA, mutations, translocations, deletions. But DNA when it does something, it gets converted to RNA, and that's captured by gene expression profiling, which measures abnormalities in the function of the genes. It answers the question, what is the cell trying to do? Some RNAs turn into protein and proteins can be detected in a variety of ways. Immunohistochemistry, which is what your pathologist is doing when they're scanning for SOX-10 or Melan-A on your melanocytic biopsies. But proteins can even be detected in the blood, so can RNA and so can DNA with liquid biopsies now.

So what's a typical clinical workflow of a pigmented lesion? Just to give you an idea of what tech is out there. So you've got a funny brown spot. Well, we used to be able to tape strip it, but the company that had that has gone bankrupt. So it's probably not clinically available for much longer unless someone picks it up. But maybe something like millimeter wave imaging will replace molecular testing in the role of answering the question biopsy or not, just the way our dermatoscope has. For me personally, I use my dermatoscope to exclude a biopsy more often than to rule it out. You pull it out and you go, "Oh, that's why it looks that way. It's not melanoma."

But let's say that you do a biopsy of a funny brown spot and it comes back as an atypical melanocytic proliferation. Your pathologist won't call it melanoma, and they won't call it benign. So what the heck do you want me to do with it? Well, there's a gene expression profile test that could be helpful for that. It's a diagnostic GEP that'll answer the question, is it melanoma or not? So if it comes back as melanoma, now what? Well, we know how to treat melanoma in dermatology, but there's some other things that can be done. We can get additional information from gene expression profiling that can give us an idea about that melanoma's prognosis for the individual patient. No longer are we making population-based estimates based on thousands of patients with a melanoma like mine. We're able to tell the patient specifically, how are you likely to do with your melanoma? A similar test is available for SCC, which we won't talk about today.

But then what? So it's melanoma. Well, what happens if the patient has a metastasis? What happens if they need an advanced therapy? Is there a way for us to monitor their immune checkpoint inhibitor? Is there a way for us to predict which immune checkpoint inhibitor they're more or less likely to respond to? Well, we have liquid biopsies for that, and liquid biopsies are an exciting space in medicine. We'll talk about both of these just briefly, the prognostic GEP, which is a clinically available test, and it turns out in a very large real-world-based study of prospectively tested patients with this gene expression profile test, if your patient gets it versus a matched melanoma that doesn't, they're 29% less likely to die from melanoma. What the heck is that about? How does a prognostic intervention, not a surgery, not a therapy, not even an imaging test, improve survival? Well, there was a really great paper that came out that sought to dissect out why patients who get a prognostic test do better than those that don't.

So in this study, they looked at patients that were tested with gene expression profiling who did not have metastatic melanoma. These patients had Stage 2B or lower. There was no melanoma in their nodes, no melanoma distant within the body. But if their gene expression profile came back as high-risk, they went on an experimental imaging protocol, which involved getting CT scans, Q six months, so on and so forth. And it turned out, if you get routine imaging, guess what? Big surprise, you find your melanoma earlier before compared to patients that don't get routine imaging. And on top of that, the tumor burden was less in patients that had GEP testing and routine scans. And this matters because the strongest predicted response to therapy is in fact tumor burden at the time of diagnosis and earlier detection of recurrence.

Once the patient's been diagnosed with melanoma or metastatic melanoma, you can check things like circulating tumor cells, which actually correlate with prognosis, more tumor cells in your blood, worse outcome. But you can also track those CTCs. And when we put someone on a checkpoint inhibitor, if they go down, that's a good prognostic factor. You can check for c-KIT mutations in the DNA that's circulating from the melanoma in the blood. You can quantify circulating tumor DNA and use that as an early marker of recurrence. You can also characterize exosomes, which you're hearing a lot about in the rejuvenation space, but it's relevant to medical dermatology too, little packages of cellular material that are circulating the blood released, not just from melanomas, but all kinds of cancers and all kinds of inflammatory diseases as well, as well as microRNAs, which we won't talk about today. But there's these really cool little RNAs that do functions in the cells independent of becoming a protein that correlate with different prognostic factors.

Speaker 1:

It sounds like we are just beginning to scratch the surface of how new technologies may shape patient care. There's been some concern about the effects of dupilumab treatment on growth rates in very young children. Dermatologist Dr. David E. Cohen spoke about emerging data at the latest Biologic & Small Molecule CME Bootcamp in March. We caught up with him on site to make sense of the data for this episode's Dermalorian Derm Decoder.

Dr. David E. Cohen:

Today at the bootcamp, we had a great opportunity to discuss in depth the use of systemic drugs in atopic dermatitis. One of the important issues we covered is the use of dupilumab in very young children. Emerging evidence suggests that kids who are small for their age that have atopic dermatitis may have the opportunity to catch up in their growth once they start dupilumab. This may be related to the reduction of inflammation, that's unclear. But we do see this emerging data that growth catch-up is possible in these vulnerable kids.

Speaker 1:

If you want to learn more about biologics, join us at the next Biologic & Small Molecule CME Bootcamp on June 7th in Scottsdale, Arizona, or hear from Dr. Cohen, Dr. Cotter, Dr. Baldwin, and more at the Derm 2025 CME Conference, July 23rd to 26th at the Encore at Wynn in Las Vegas. The agenda for the four-day education packed conference has been published. Get information on both events at dermnppa.org. Thanks for joining us for The Dermalorian Podcast from the Dermatology Education Foundation. The Dermalorian Podcast is produced for DEF by Physician Resources.