Episode 5: Artificial Intelligence and Deep Learning

Artificial intelligence is a rising force within dermatology, with increased scrutiny and application over recent years. What does the current evidence say about its efficacy? Where should our next research priorities lie? And will we all be replaced by robots by the end of the year?! To answer these important questions and more, Dr Jonny Guckian chats with Dr Beibei Du-Harpur, Doctoral Clinical Fellow at the Francis Crick Institute.

Transcript

SPEAKER_01 0:18

Hi there and welcome to BJD Talks, the official podcast of the British Journal of Dermatology. In this podcast, we look well beyond published studies and try to dive into the real-life implications of dermatology scholarship in what's a relaxed, accessible way. This podcast should be helpful for anyone interested in skin health research, whether you're a dermatology professor, researcher, registrar, patient, or simply someone with skin enthusiasm. We hope you'll join us as we build in our world leading research through friendly discussion. My name is Dr. Johnny Guccian, and I'm a dermatology registrar in West Yorkshire, as well as the BJD's Podcast Associate Editor. Together we'll explore issues as wide-ranging as patient and public involvement in research, global dermatology, and social media and dermatology research. Technology is influencing the way we practice dermatology. More than perhaps any other area, artificial intelligence represents an exciting and promising field for both research and practice. But where's the evidence? Is AI really going to change our diagnosis and management of patients? And the big question, are we all going to be replaced by robots? To answer these questions and more, I am joined by a brilliant guest, Dr. Baby DeHarper. Dr. De Harper is a dermatology registrar and clinical research fellow at King's College London. Thanks for joining us, baby. Thank you very much for having me. No problem at all. So AI can be a minefield of complex information some clinicians and other staff struggle to conceptualize. Tell me, how did you get into AI?

SPEAKER_00 1:49

I started my PhD in 2018 and I joined a computational biology lab based at the Francis Crick Institute. And I'm co-supervised by Fiona Watt and Magnus Lynch at KCL. And Magnus has a really strong interest as well in AI. He's also a dermatology consultant. And I thought this was a great opportunity, given the people I was surrounded by, to learn about AI and its applications in medicine, especially dermatology. So it was more just that I had an opportunity to learn about it from people who had relevant expertise. And the lab I'm I'm based in at the Crick, they're all bioinformaticians, computational biologists, people who are, you know, designing convolutional neural networks. So there were people in the lab who had expertise, and I think a lot of people really think about PhDs as being a project, but it's also a project on yourself and sort of upgrading your own knowledge and your skills. So that's kind of how I got into it. And I started purely by chance doing a bit of work with a company in London called Skin Analytics, who are a dermatology AI company. So I also learnt a lot through working with them as well. They work with lots of different dermatologists. So yeah, that's kind of how I got into it, and I ended up writing a review paper which was published in VJD, which I think lots of people have read and you know from conversations, and I think it was hopefully quite useful for people. So yeah, that's kind of my background.

SPEAKER_01 3:22

That's great, and yes, I can confirm my department used your review paper as a journal club article at one at one point a few months ago, and it was well received. So there's some there's some good feedback. Um you've you've mentioned quite a few interesting, uh slightly complicated terms, including the various aspects of AI, and I know in the review it talks about things like machine learning as well. There's a lot to take in um here. What is your elevator pitch to dermatologists for what artificial intelligence really means?

SPEAKER_00 3:55

So my elevator pitch to dermatologists is have you ever sat in the middle of a busy skin cancer clinic and thought to yourself, I wish there was more than one of me? We all know as a dermatologist that visual pattern recognition is a really important skill, and we spend a lot of our training sort of building up our ability to recognise patterns in the skin. And AI, and I say that in inverted commas, feeds large quantities of data into a self-tuning algorithm, which uses a lot of computer power, to perform relatively simple pattern recognition tasks that is designed to mimic the dermatologist. So I think that deploying AI has the potential to scale up the access to a dermatologist's pattern recognition skill. And I think that this could be of benefit to patients in terms of access and also of benefit to the service we're able to provide as dermatologists because we know that our waiting lists are getting longer because the number of patients coming through skin cancer clinics are just growing and growing year on year. And it's unsustainable. I don't think it's right that we can't see children with eczema in a timely manner. I don't think it's right that patients with severe inflammatory skin diseases are perhaps not as prioritised. Obviously, skin cancer is very important, but we also know that the patients that do come through our door don't all have skin cancer. I mean, a lot of them do, but a lot of them also don't. So if there's any way in which that workload can be improved so that we are seeing the patients who need to be seen, then I think that's only a good thing. So that was my conclusion from um, you know, talking to various people in research and industry about how AI could be beneficial. And yeah, like I said, I think of AI as in inverted commas because I actually don't really like the term. You know, we think of intelligence as being creativity and being able to think on your feet and all of these other very human skills. But in reality, what we are calling AI, which is probably like the sexiest term for it, is it is simply algorithms learning from data. And these algorithms are able to do this because people have become better at designing these algorithms, and probably the most common one used is something called a convolutional neural network with many, many, many layers. And that's called deep learning. That's the most common, but you know, they're just performing actually quite a simple task, and the scope of what they can do is quite limited because they are learning from data. So, what data goes into the algorithm is critical to how the algorithm functions, and I think it's important for people to recognise that.

SPEAKER_01 6:28

Yeah, and I've I've heard people working in AI talk a lot about this about the importance of the human end in inputting data and correct and you know, selecting the right data and interpreting data. And I think those who see AI as this catch-all solution to everything that is just a magic machine that's going to solve all of our problems kind of kind of miss that. It's going to require human input and human error, I guess, um, as well as everything else does.

SPEAKER_00 6:53

Absolutely.

SPEAKER_01 6:54

I think that's that that's really cool. Um just just kind of hearing you you talk about that this is there's a lot of really um exciting stuff there. The concept of deep learning is something I've found interesting and just was wanting to get kind of get my head my head around that. Can you just explain that? Can break it down for us a little bit, deep learning.

SPEAKER_00 7:11

Gosh, I'll do my best. So, deep learning is uh often something that's referred to as part of say a convolutional neural network. So it's really about how information is passed through multiple layers, and deep learning simply refers to a network which has many, many, many layers. So the difficulty is with these networks is that we don't know what's going on within these layers because they're a self-tuning algorithm. So part of how they work is that they have this internal sort of feedback mechanism where they tweak the functions in each node so that the performance becomes closer and closer to the gold standard that you've set. So if that gold standard is, say, a histopathological diagnosis, then it will tweak the internal functions within each node until it is hitting that the standard that you set. Does that sort of make sense?

SPEAKER_01 8:03

That's great. And we will we will yeah, that I mean that that that explained it beautifully, thank you. And we will talk a little about those standards and later on and whether the AI at the moment is meet is meeting those standards. Obviously, you've worked quite a bit and you've written about um AI and you work with others who've worked in AI as well. What excites you about artificial intelligence again in a vertocombus?

SPEAKER_00 8:23

I think that um, you know, like I said earlier, it's about the opportunity to provide better care. I did quite a lot of stuff in QI in the past, quality improvement. I'm sure you know lots of other junior doctors have done that sort of thing. It was kind of mandatory as part of the foundation programme, but I really got quite into it and I love problem solving. And I think, you know, as a derm reg, you see the problem of the two-week wait. I'm sure you've seen that as well. And how can we innovate as dermatologists, given that you know, we know that the workforce is an issue in the UK, you know, especially like being on social media as well. Like people talk about how they can't see a dermatologist quite a lot, and you know, obviously you feel bad that this is the situation. I'm not saying that AI is necessarily the answer to all of these issues, but I think that we just need to think creatively about how we can solve problems and view advanced technologies as providing part of the solution rather than just with suspicion. Um, and I think that people often think of these things as being a threat to their livelihood, and I really don't think that would be the case. When ECGs were invented, cardiologists didn't think, right, I'm out of a job. I don't think it's really a concern.

SPEAKER_01 9:40

That's a really good, um, a really good analogy. In fact, I think this cardiologist got a lot more work, really. Not electrophysiologist, then so yeah, I I love that. I think basing what can initially seem quite a cold and and uh intimidating concept, uh basing that around real human problems and opportunities and values is really impressive and it I guess it can help drive our um needs based on so these QI principles, for example, and trying to solve problems um from a problem-based approach rather than a solution necessarily based approach. I think I think that's that's really cool. And um again, yeah, um applying an open mind to newer technologies requires a bit of a cultural shift, and we might chat a bit about that um later on um about how we integrate this better into our our systems. So we've been chatting about the fact that AI can solve problems, but let's talk real world. What are these applications of artificial intelligence? What are we doing at the minute with AI and what could we do in the future with it?

SPEAKER_00 10:43

I'm more familiar with what's going on in the UK than um in the US, for example, so I can only speak of the UK really. But I mean some of these companies are international, just to mention. I guess there's kind of two approaches. There's the direct-to-consumer approach, and then there's the physician tool approach. There are a couple of companies, including Google, who are going for a more direct-to-consumer approach. So Google haven't launched their product, but they've done a little preview of their dermatology product, but um, you know, is able to recognise I think over 200 skin diseases. I'll have to check the number exactly, but you know, it's a lot of skin diseases. I think it's basically like a Google search, but with a picture, essentially. That's the sort of my broad understanding of what they're able to do. But obviously, it's exactly that. So rather than typing in, I have a rash on my foot, you know, they can take a photo of their foot. I think that's what they're aiming for. And then you have um a couple of companies doing stuff in terms of uh skin lesion recognition um to try and help people um identify skin cancers. So there's a company called Skin Vision, which is probably the biggest direct-to-consumer app, and then there's Skin Analytics, um, which has done some NHS deployments as a sort of interface between primary and secondary care.

SPEAKER_01 12:02

Brilliant. And am I correct to say that most of the work that's been done in this particular area at the moment is with regards to identification of skin cancer? Yeah or is there a more broad approach?

SPEAKER_00 12:12

So the Google application is not just skin cancer, it's all kinds of things. But their diagnostic accuracy is quite low. So I think that is more meant to be like uh something that people can do. I'm not not saying for fun, but just as a as a sort of adjunct to maybe a general Google search. That's my impression. But the technology in terms of skin cancer recognition is a lot easier because the nature of um you know skin lesions being more limited because the size, the image capture methods, you know, we have lots of image banks of clinical and dermoscopic images, um, which you can feed into machine learning algorithms. So that's why that technology is a lot more advanced. I think realistically, you're quite far off doing more with inflammatory skin diseases because they're they're so variable in their presentation anyway. I think there's more of an art to it, um I guess, which is the more human side, uh, rather than a simple visual pattern recognition thing, which is really what machine learning is good at.

SPEAKER_01 13:16

So I guess for the time being, we're still going to get plenty of um messages through on our social media um DMs from friends and relatives asking us what what this rash is and what that rash rash rash is. It's not going out going away anytime soon. They're not turning to Google just yet.

SPEAKER_00 13:31

Yeah, well, I mean we'll see how the Google thing works. I'm sure people will find it interesting. Um, I mean, I'm I I'd love to test it out once it's uh once it's launched.

SPEAKER_01 13:40

Great, yeah. And I I would say other big tech companies are available. We have no affiliation with Google's AI, um, but it's uh it's just it's uh absolutely fascinating to see what's happening in kind of the within the industry. Um and you know, we talked about need earlier, there is the systemic need within the NHS, but there is also the patient need, and as we've talked about in previous episodes of this podcast with patient and public involvement, patients should be driving a lot of what we do, if not most of what we do, because it's society's needs that we serve. Okay, so you talked about efficacy earlier, and so I guess the next question is does all of this work? What is the state of the evidence base at the moment and for AI and dermatology?

SPEAKER_00 14:23

There was a nature paper published by Esteba et al. uh several years ago now, which I'm sure you might have come across, and that was probably the first big dermatologists versus AI paper that came out. And there have been multiple similar papers that come out year after year showing that an AI can perform similar to a dermatologist. But you know, the reality is that what they're doing is they're they're they're testing the dermatologists in the machine learning slash AI's home territory, which is images that are just sort of presented to someone on a screen. So it's nothing like the process of actually examining a patient. We don't understand how convolutional neural networks work exactly. So it's a bit of a black box. We don't know what they're picking up inside an image that's making them say this is a cancer, this isn't a cancer. So, you know, the the algorithm is a bit invisible to us in a way. So, for example, you know, when they're learning from a data set, there may be things that are unique to that data set which they are kind of incorporating into the algorithm, because that is their world. So when this algorithm is being um sort of self-tuning itself, it's taking in all this data. And if that data is limited in its scope or has like a weird characteristic about it, for example, surgical pen markings, that can fool the algorithm into thinking this surgical pen marking is something meaningful in some way. So there's always this uncertainty about, you know, when it when you go from you know learning from this limited data set to going to a completely novel, different data set or slash real world, whether the algorithm will still perform to the same um standard as when it was trained on the original set of data. So it will always perform really, really well if you're testing it on images that came from the same data set. Like we know that in general. But when you then test it on images outside of its original data set, then a lot of the time these algorithms perform worse. And I think this is especially a problem with algorithms that have been built using a method called transfer learning, where they have a pre-trained network. Um so Google have one, I think Facebook have one as well. And these pre-trained networks uh apparently are quite prone to this issue where they're not able to extrapolate suitably to real-world settings.

SPEAKER_01 16:53

Yeah, and I guess that's the Rob, isn't it? It's how it performs in the wild. And I I I was at a talk recently which just kind of describes tech well technology more generally as a drug, and it's like when you have the evidence base for any drug, it's in its very, you know, isolated environment that is in very strict test conditions, but then you bring it out into the real world with real people, real humans with comorbidities, and you know, and maybe they are fallible in terms of taking said drug, etc., or have got multiple other medications, their interactions doesn't always work as well. Um so if you appreciate AI in the same context, it's not going to be as perfect as you will get in the in the original papers, will it?

SPEAKER_00 17:31

No, oh, absolutely not. And I think that if we want to use AI as a diagnostic aid or diagnostic tool, then it should really be evaluated in a real-world setting. And I think that any deployment should be done alongside engagement with clinicians, educated clinicians, and active research. Because I think it's important that we recognise that this is something that's here to stay, and it's better that we take ownership of the problem/slash solution, because it is both, and and just engage with it and make it work for us in the best way possible.

SPEAKER_01 18:08

Yeah, I think it's having a accepting open um environment that is lightly controlled initially and you know it can ease its way into it. When you don't do that properly with AI, it can go a bit haywire. I know there was a case, or there'd been a few cases, of I think Twitter or similar similar big tech companies have created AI bots which learn from the Twitter environment and soon became xenophobic Nazis who decided to swear all over the place because that's what Twitter taught it to do. Um which I don't know what that says a lot about the human about about humanity and the human psyche. That's terrible. Yeah, so yeah, um making sure we keep things in safe environments before we uh um we unleash it into the wild. And I think your your description of easing it in alongside dermatologists um is is a is a really um helpful one. And the idea of dermatologists versus AI is just fascinating. It sounds like the sequel to God's delivers a column. Um But I mean it is interesting because the AI is winning in the papers that you see published. And if you look at that in isolation, and I guess if you look at this is something else that came up, came up in a talk that was at recently, you look at those who are delivering services, if you are designing services, maybe regulators or maybe maybe those who are committed commissioning services, they will look at the data and they will think, oh, this is great. We have a workforce crisis, we have a a um a need for more dermatologists, but we don't have any more dermatologists, so why can't this do it now and why can't it it do it now? So I imagine there will, in terms of a service provision kind of conundrum, there will be there will be attention there.

SPEAKER_00 19:43

Definitely. But I think we have to think about realistically when you deploy an AI, what's going to happen? Does it mean that you're going to biopsy more patients? Does it mean that you actually get more referrals because people are just self-diagnosing in some way and then presenting to GPs and saying, Oh, my app told me I've got skin cancer? We need to have data about what is actually going to happen in a real-world setting to really try and anticipate the impact on a healthcare service because I don't think it's a simple solution. It's a diagnostic aid, it's not a consultation. Like the number of patients who come in to a skin cancer clinic, um, they come in for reassurance, and um, sometimes they're coming in for a full skin check as well. You know, that there's lots of things that are going on in a consultation that is not captured by a photo and a letter in response. I think that's not looking at the bigger picture, that's not looking at patient satisfaction either. So I think we've got to look at the whole situation quite holistically, and I think also it's only doctors and dermatologists who are going to understand that those are the issues in a skin cancer clinic. So I think it's important to try and you know engage with the process and be as well informed as you can be so that you can have you know reasonable conversations about what the limitations are, what the risks are, and all of those kinds of things.

SPEAKER_01 21:08

That holistic approach is is really important because often even just reassurance has value in itself because it's giving an explan a clear explanation and it is you are introducing uh an intervention there by providing health information, and you're you are preventing skin cancer in the future, potentially by encouraging monitoring, for example. So so if we're talking about adding value to care and adding value to patient interventions, then you need the human. You still need the human. Absolutely. A lot of this seems a bit abstract because we're talking about these fancy algorithms and how it might come into um our practice specifically to skin cancer care, but a lot of us here in the NHS, and maybe even in the world beyond the NHS, because apparently there is one, um are struggling with slow IT and you know computers that might take half an hour to turn on. And just and that's not that's even without logging in. What are the barriers to integration of AI into our systems that that you're aware of? And how far away do you think some of this is?

SPEAKER_00 22:10

There is already the existence of some integration because I know that, for example, skin analytics is part of some care pathways. I think a big barrier is the cost, actually, the cost of doing research in AI and dermatology. So, you know, we obviously, as evidence-based clinicians, we like to see that something is well validated, well studied before we like to incorporate it into our clinical practice. So the same way that you know we are suspicious of a new drug, we should be suspicious of new technology. But you know, clinical trials for drugs are expensive, and clinical trials to show a diagnostic tool does what it claims to do is also expensive. That definitely is a limiting factor. And these trials can take a long time as well. So there are many things that play a role in delaying the deployment of an algorithm into the clinical workspace.

SPEAKER_01 23:04

Thank you. So that's if you've got trials coming through, or we've got um work that's that's being done out there that that sounds exciting. Let's for our just for our listeners, if they were to go and maybe look at different papers out there that that might you know show the the work that's being done right now, are there any direction you you would point them in in any particular papers that you'd recommend um our listeners have a little look at?

SPEAKER_00 23:26

Well, so my review paper summarises quite a lot of the uh sort of the studies um that have been done. So the vast, vast, vast majority have been done in uh not a real world setting. So um, in terms of in a real-world setting, there's a paper published in uh JAMA Network Open, which uh was a study that was conducted at the World Free. So that one is worth having a look at. I can't remember the authors off the top of my head, I'm afraid. If you want to look at a prospective study, I think obviously there's loads and loads of aspects that we need to look at before you know maybe we'll feel more confident. You know, the same way that you know, if there's a new psoriasis drug that comes onto the market, there will always be people who are more early adopters and some people who will be more hesitant. And I think you might be an early adopter if you've read more about the drug, um, read the clinical trials yourself, and you've done a lot of sort of severe psoriasis clinics, so you're more comfortable with that kind of prescribing. So I think that um developing your own um sort of confidence in the area, I think, will help more dermatologists feel confident in adopting the technology if they feel like it would benefit, and also you know, recognising the limitations. I really think that education is key, not just for dermatologists, but also for the patients and the public as well, because one of the issues with the term, like I said, artificial intelligence is that it implies intelligence, and I don't think it is intelligent. I think machine learning is a lot more useful as a term, personally. And I think that if you if you understand the limitations, you you recognise that it really can't do certain things, and then you won't necessarily try and use it inappropriately. I think that's really key. And also for patients not to use it inappropriately as well.

SPEAKER_01 25:15

Yeah, great, thank you. And for our listeners, I really would recommend uh Dr. De Harper's uh review paper because it it does break down some of the concepts that we've been talking about and touches on some of the uh applications and the future directions as well and sums up the literature quite nicely. So, as regular listeners of this podcast might know we do like a shameless plug for our our um guests' work, so we're more than happy to more than happy to uh recommend um that review. You mentioned education, and some of the the researchers and other academics that I've spoken to within AI in the past have highlighted staff education and patient education as one of the biggest challenges uh or hurdles that we have to overcome if we're going to successfully implement uh machine learning or AI or um or however however they may term it. How do we then educate our staff as to the benefits and risks?

SPEAKER_00 26:12

I think it's a difficult one. I think that definitely, for example, as a registrar, you'd want to maybe have a teaching session on it. Um, you know, I'm a dermatologist. I'm not someone who's designing an algorithm from scratch. So I think it's useful to have people who have specific expertise that's different from your own who come and talk to you, you know, regulatory experts, things like that. So just having breadth of exposure to people with specific domain expertise and being educated from them, I think is really, really useful. And that's certainly what I found most useful in my own learning process.

SPEAKER_01 26:48

Yeah, and and I guess as with most educational interventions, some of that has to be systemic and cultural. It has to be include staff who are open to change in new things and new ideas, and open to, as you say, industry coming in. We have that in with drug reps, you know, I mean heavily regulated, of course, but come in and do give talks on new medications, and that's how these things, the new therapies, get started. So perhaps they need to have a similar approach with AI um or similar technologies. Who knows? So, yeah, there's there's has to be a bit of a forward-thinking and enabling approach. And I guess if you try to teach the registrars now, kind of our cohort, and we'll see you know that that change moving throughout the generations of as as we progress and become consultants uh in the future.

SPEAKER_00 27:32

Definitely.

SPEAKER_01 27:33

Yeah, it's it's about being strategic, I guess. And it would be important to point out now that that the BAD, uh the British Association of Dermatologists, have a an artificial intelligence uh working party group um as well, who are doing um some strategic work in terms of AI. I know they've got some plans, particularly with regards to education, so you can check out uh their website or their section of the BAD website because there's a lot of interesting reading there. Well, this has been absolutely fascinating. I know I'm I'm learning a lot and um it's complemented the review quite uh quite quite nicely hearing some of the um real life applications and the challenges that we may encounter. And you know, putting the uh well, questioning the intelligence and artificial intelligence is is uh is a really interesting take on all of this. That that brings us to an end of today's episode of BJD Talks. We've hopefully demystified AI a little bit and highlighted some important reading and reassured our listeners that we will likely still have jobs when the robots take control. We look forward to sharing our next episode of BJD Talks. In the meantime, do please let us know if there are any hot topics in dermatology you think we should discuss. Of course, we're on social media, so please try to reach out to us uh via at brjdermatol on Twitter and at brjdermatology on Instagram or by using the hashtag BJDTalks. Dr. Baby Day Harper, thank you so much for taking the time to come and chat to us.

SPEAKER_00 28:55

Not at all. It was a pleasure. Thank you for having me.

SPEAKER_01 28:57

No problem. Um, bye for now.