Skin Cancer AI

I grew up in Australia, and skin cancer is kind of our national cancer — we're almost proud of it. The incidence of melanoma is very, very high, around 40 per 100,000; we've got some parts of Australia that are over 70 patients per 100,000 who end up with melanoma. So we're really good at making it, to be honest.

I grew up with all the public health messages that were out in the 80s. There are some really great campaigns that every once in a while we dig out and share with people — "Slip, Slop, Slap" in the sunshine. So it's always been in the back of my mind. And then — I generally don't talk about these sorts of things — but in my family there have been some health problems in and around cancer, and then a friend of mine. So it's always been in the back of my mind that I wanted to do something that could make a difference and that mattered to me. And it felt like being able to solve a problem that disproportionately affects younger people really mattered.

So when the opportunity came along to say, okay, I want to try and do something in healthcare — because that's what I really like — I asked myself: what do I know? Well, I've got a maths and physics background. I kind of get this AI side of things that's emerging. And I really feel like there's some benefit in using computers to create capacity — to do some of the things that we could only do with humans before. And then it just snowballed from there.

But we took some really important decisions early on that have coloured us as a company. They weren't necessarily as popular then. They seem like no-brainers now, and everyone talks in this language, but at the time they were considered a little bit conservative, or maybe not the right strategy.

Firstly, we said we're only going to work with high-quality data. We're going to pair histopathology outcomes with the images we have to work with. That means that instead of having hundreds of thousands of images, we're going to have a fraction of that — and we're going to have to figure out all the technical challenges of applying artificial intelligence to that so the technology still works. That's actually where a lot of our IP is: scaling the networks and building from the ground up so they're designed to solve the problem we have. It's different to tell the difference between a malignant lesion and a benign lesion than it is to tell the difference between a traffic sign and a beach — the network architecture is different. So that was a key decision: we're willing to take the technical risk and overcome it, rather than say, okay, we have to work with as much data as we can get because that's just how AI works.

The second one was, we decided early on: we're not an app. We're not going to be distributed through app networks. We're not just a glitzy consumer toy. We're something that fits within the clinical pathway, and we want to take the responsibility of being part of the clinical journey for a patient. At the time we made that decision, everyone was going out on the app store — everything was an app. The feedback I got was, how could you plan a distribution strategy that doesn't give you the entire world immediately? You're making a crazy set of decisions. But it allowed us to say, we want to be part of the clinical pathway, and we want to use a dermatoscope to get these higher-quality images that make us better at the diagnostic aspect of it. All of that strategy tied together once you're not considering yourself an app.

The third and final thing — which tied into the app strategy — was: we want to be a clinical product. And if you're a medical device, there's no way around it: you need to lean into the medical-device regulations and adopt them, not just see them as a box you tick to get approval to sell. You have to embed it into the company from the ground up. So we've got a regulatory director in our business who's fantastic, and who's been tasked with: help us meet these regulations, but also take all the really good stuff out of the processes you need to build to meet them, and figure out how to make that a competitive advantage — so we can build high-quality products faster than our competitors, because we really lean into it.

I think those three decisions were core to who we are as a business. We made them years ago, and they're now kind of accepted as what you should do. But at the time, that was our secret sauce.

20% data, 80% algorithm. I personally feel there's a massive misconception that AI is a data game. Clearly the data is important — as it is in most things, when you're making decisions, more data means better decisions. But the reality is that the more data you get, and the more understanding you get about the AI, and the more chance you have to optimise and really build out something that helps clinically — the more complex it is.

I make a terrible analogy about golf, and I need to get a better one. Everyone knows golf is one of those really frustrating games — you can pick up a golf club and hit the thing around, and get to a level where you have a bit of fun. I actually don't play golf, so I'm guessing. But if you want to be any good at it, you've got to spend a lifetime working on it. That's where we are with AI.

And that may not matter if you're suggesting the next film someone might like to watch. But when you're making clinical decisions, it does matter. A small change in performance makes a huge difference — especially when you think about what part of the operating curve AI works at. You're always trying to optimise sensitivity and specificity, and you're working at the very top range of sensitivity, which means a very small change has a very big impact on your specificity, or your ability to discharge patients. So the performance is all-important, and that takes a lot of work. I'm a bit biased, but I think it's 80-20 to the algorithms.

Yeah, absolutely. And you know what — on the more human side of things, one thing that clinicians are trained to deal with, that those of us in health tech who aren't clinicians haven't been, is that you make decisions about patients that end in bad outcomes. Sometimes it's a decision you can track back to, that you could have made better; and sometimes it's just outside your control. But you know there are going to be patients with bad outcomes, and you're trained to deal with that.

We had to have a really open, honest conversation amongst the team — especially our head of AI, Jack — to say: we're building something that is going to make wrong decisions, and a patient is going to have a bad outcome. And that's going to be on us, because it's our technology. We took that really seriously, and did a lot of soul-searching about it. Where we got to is: do we feel like we're actually making the pathway better, the experience better, so that the net effect is a positive outcome — versus the single person who gets a worse outcome because of our system? It's not an easy thing to deal with. It's something we've had to train ourselves around, and to make sure the mental health of our team is looked after as they consider their role in building a system that's going to have that sort of outcome.

So putting that human bit aside first: yes, it does keep me up. You're trading off your ability to find a disease — cancer — versus your ability to deliver value to a health system that's overworked. The way we deal with this is that we have an external clinical advisory committee: dermatologists internationally and within the UK, a public health expert, and a health economics expert. We get together as a group and ask: what's the right trade-off? Are we making the right trade-offs? Are we approaching this the right way?

The other thing we do is that when we built the medical device, we built it so you can configure the sensitivity/specificity trade-off as part of the device. So we go in with our partners and say: we know that a dermatologist typically, in the literature, finds 92% of the melanoma they see first time. What sensitivity do we want for melanoma? And we can configure that. Then: what sensitivity do we want for squamous cell carcinoma, and basal cell carcinoma? We work our way down the list and configure the algorithm to do that. Then we can have an informed discussion with our partners: this is what the potential discharge benefit looks like. Is that sufficient? Have we got the right trade-off? So we very much see ourselves as partners with the clinical teams we work with, rather than prescriptively saying, this is what you're going to get. But we always want it to be above 95% on all lesion types — and that's generally above what the literature suggests you can expect across a pathway for clinical teams.

So — I guess that's a tricky question, and we get asked it a lot. I don't think it's definitive in terms of the answer. But the principle that matters most, and where we start, is that at Skin Analytics we want to be involved in the clinical pathway. We're willing to acknowledge that that means we wear responsibility for the decisions that we take and help people get to.

Each of our pathways is designed with our partners, and the AI plays slightly different roles. Sometimes it's more supporting a decision from clinicians who are trained to make those decisions. Sometimes it's more autonomous, in that it makes a decision with a patient to discharge them that may not be reviewed by a dermatologist — and we can only do that because of the huge amount of performance data we have. But at the end of the day, we're not shying away from our responsibility and liability within a pathway. When you start from that principle, you have pretty productive conversations with your partners.

The reality is, when it comes to healthcare and you get to the discussion around liability, it almost doesn't matter what people say about it. Each and every person involved in that pathway — their decision has to be deemed reasonable by a group of outside experts. If it ever got to the point of asking "was the right decision made?", the outside experts look at what happened and what information was available at each point along that journey. Was the right decision made? I think that's a really healthy place to be, and we're quite comfortable working within it.

So it's a long way of saying: it varies by partner. But we do take liability, and we invest heavily in clinical safety — another one of our management team is our clinical safety officer. We're always thinking about the ways things can go wrong for patients, and putting mitigations in place to minimise that, just as you do in any clinical service you design.

As with every question you've asked me, I've got a little side note first. One of the authors on that paper is a guy called Justin Ko, who's head of dermatology at Stanford. I got in touch with him after he wrote the paper and said, hey, really amazing work, this is what we do, this is how we're approaching it — and we just got on like a house on fire. So he's actually one of our clinical advisory committee members and one of our advisors, and helps us think through how you take the principle they showed in that paper and apply it to the real world.

One of the challenges with that paper was that the approach they took doesn't translate that well to the real world. And I'm going to paraphrase my head of AI, Jack, who's probably going to be mortified at what I say here — but you have these networks that are really large, and you need these huge transfer-learning data sets to train the algorithm. Then you give them something like skin cancer, where you've got relatively few degrees of freedom between a benign and a malignant lesion — compared to a tree and a stop sign. So now you've got networks that are just way too large for the problem you're trying to solve, and that creates all sorts of issues.

We do actually use transfer learning — it's not that we don't. It's just that we didn't start with the Google Inception network. Well, we actually did start with that years ago, and the preceding ones — the one from Oxford and a few others that we used to bundle together. But what we realised is that that approach of using these other networks — they're just not designed for the same problem. So when we say we built it from the ground up, we changed the network architecture. We still use a lot of things like transfer learning. In fact, I think we've got a blog article where we talk about ice creams and kebabs being a key part of us helping to find skin cancer — because a lot of the data that seems nonsensical is needed to pre-train these networks. We as humans have grown up collecting information that helps us understand what the world looks like, and then we're asking these AI systems to come from nowhere. You need to give them some of those background parameters as well.

And a little funny story on this. We were in a co-working space back when we were doing it. When we started doing some of this transfer learning, everyone was using ImageNet — this massive collection of labelled images that started at Stanford and Princeton, or something like that. But in the GDPR days, we weren't sure the proprietary rights were given for those images to be used, and we didn't want to use them in our training. So we said, okay, we're going to find data sets we can use. We joined things like Flickr and a bunch of other places where you could get the right licences, and started downloading huge amounts of data. And then we got the entire co-working office banned from a bunch of sites we were collecting data from — because we hadn't done anything wrong, but their automated tools picked up these huge downloads and blocked everyone for about three or four months. So we were hugely unpopular for a while.

The limitations come in when — yes, they're fantastic, incredible algorithms, the data does exist, and reinventing the wheel doesn't make sense. But what you see is that there's a very large and recognised drop in performance from the lab into the real world for many AI systems in healthcare. And that performance drop, as we talked about earlier, is just critical in healthcare. Dropping 10, 20, 30% of performance is catastrophic for care, because you have expectations of what that care needs to be.

Where that drop comes from, in our opinion — and I don't think it's quite this simple — is that when you have a huge network and you don't have a huge data set, so the clinical image data set you have is not as significant, these networks are so big they almost memorise the data you've got. Then you're optimising a system for the data you've got to train it, rather than for the real-world data. So it starts learning the wrong things.

That's why we've had such great success building our networks — because we were able to balance out all the different inputs and constraints we have, which are actually different to something like the Google Inception network. Because there's just so much data for them. The output of our algorithm is really resilient: what we saw in our observational paper, what we saw in our prospective study, and then what we saw in the real world are actually pretty much the same thing. And that, I think, is because of the network design.

I think you hit the nail on the head on the different ways you can use it. Where we want to get to in the long term is: we have a capacity constraint in clinicians. Someone said this to me the other day and I really like it — it takes 15 years to make a specialist, but it takes nine months to make a patient. So we have this disconnect between the ability to make very difficult decisions around care and the demand coming through. In dermatology, we're 25% understaffed in the UK; it's up to 50% in the US. There just aren't enough specialists.

So what we want to do is make it so easy to get a skin cancer assessment that everyone's doing it all the time, we find the cancers at an earlier stage, and the survival rate's almost perfect. That's where we want to get to. But as a healthcare system — and we see ourselves as part of the healthcare system — we need to ask: how do we best use our clinicians? It's always been that a clinician is the only one capable of making these diagnostic decisions — I won't quite say diagnostic, because you confirm with histopathology, but let's say that for the sake of discussion. If we build technologies like what we have at Skin Analytics that can tease apart the basics — okay, here are the patients who very likely have a malignancy, get them in to the clinicians — then the role of the clinician is actually that really complex bit that an AI system is not going to be good at: okay, these are complex cases, what is actually going on here? And on top of that, what do we do about it?

Because if you find a basal cell carcinoma in a 90-year-old patient, you're probably not going to cut the thing out — by the time it causes a problem for them, they have other issues, and do you really want them to have a large open wound to heal in their 90s? You have to weigh these things up. These are human decisions that need to be made by humans. That's the trick with medicine: it's such a human condition, a human discipline. But how much of it can we carve out to be done by technology to improve the care pathways — not to try and automate them or remove responsibilities from clinicians? That's ultimately going to fail.

So that's how we see the world of where we want to get to — but that's a longer-term journey. It's a difficult discussion to have; we need to build the confidence of the clinical community to start having it, and we're just not there yet. Where we see our role right now is: what are the tactical, real-world problems we have now that we can solve? In the UK, where we're doing quite well, the real problem is that we have a commitment to see patients with suspicion of cancer within two weeks in our hospital specialist teams. Dermatology is the highest-referring specialty. We're putting so much pressure on those teams that they're actually doing a reasonably good job of achieving it, although it's increasingly difficult with COVID — but at the cost of seeing patients who have non-life-threatening dermatology diseases.

There's a great book that's just been published called Under the Skin by a UK dermatologist. The opening of the book is this dermatologist describing how they spend their day seeing healthy people trying to find those suspicious lesions. And at the end of the day, they're rushing through seeing someone who has acne, who unfortunately has been waiting four months and is now scarring, or has psoriasis that's been driving them crazy and dramatically affecting their quality of life. We're rushing the patients who need care because we're seeing a lot of healthy patients trying to find those cancer patients. It's the right decision with the resources we have, but we've got to get to a better solution.

So that's why we started by saying: that's the problem. Can we deal with a two-week-wait capacity? Can we use the AI to triage out those patients who are clearly benign, and then focus on those more likely to have malignancies? That's our sweet spot right now. But as the business grows and evolves, you're right — we're going to start to look at new models where we can be used by a primary care doctor. That presents a challenge, because we already ask primary care doctors to do a lot, and now we're asking them to do something else. So increasingly what's popular with our NHS partners is this idea of community diagnostic centres, where skin cancer could be treated the same way STDs are with the GUM clinics — you can signpost people to ways to get an assessment outside of the typical care pathway. I think we'll see more of that as the NHS figures out how to deal with this huge outpatient recovery programme they're working on.

So — no idea is a bad idea, and we've talked about it. From our point of view, there's a way to get a lot of public attention doing that, and in many respects it's viable — the way our technology works and the performance numbers, we could do this. The challenge is that as a business, if we're going to let lay people use it, we're going to want to turn the sensitivity really high up. Because the last thing we want is to give it to someone to check something, and they go, "oh, I checked this with this technology and it's fine, I don't need to worry about it" — and they're sitting on a melanoma.

So we want to turn the sensitivity right up. And then you have this population-health question: you turn the sensitivity right up, and even if you have a low false-positive rate, you're going to flood the healthcare system and create more pressure. As I started with, we see ourselves as part of the healthcare system, trying to solve the problem for the healthcare system before solving it at a population level. So I think we'd probably not solve the problem if we tried to go about it that way.

But as a technology, what's interesting is that as we use it more, we start to see the real power of the system to be able to do that. On top of that — right now there's a single test: we look at an image, and if we have 96% sensitivity, that means about a 4% miss rate, which is good on the industry side of things. But as you pointed out, any misses are bad when you're a new technology. So we want to get to a point where we have that test, but we also have a follow-up — because we can do that pretty painlessly — and then a second independent test looking at the way things have changed. We actually have a pattern for how we do that. Then you have two independent tests, and if they both have a small error rate, their combined error rate becomes much, much smaller on the back end. When we're ready to do that and can be confident we're not going to miss things — that's the time we'll probably push more into the consumer side.

The final point — and this probably resonates with you as a clinician — is that if our job is to go out and start doing assessments in the population, we won't do that until we have a viable path to solve the problem for the patient. We think in pathways, not in parts of the pathway. I don't want to say, "hey, cool, there's something there, good luck, see you later, here's a report." We want to make sure they have a way to get to that next level of care. As a patient, you're always operating with less information than a clinician, because the clinician is trained and knows the potential outcomes. A patient with less information either worries a lot and takes immediate action, or doesn't know they need to worry and doesn't take action — and either of those outcomes is pretty stressful. So we're all in on making sure the patient has a clear path to resolution.

Yeah, absolutely. One of the things we want to get to is really being able to have an impact on the total survival rate for skin cancer. And that's not actually a healthcare problem. One of our early advisors introduced us to this concept of total patient delay. When you look at skin cancer, the delay from seeking help to getting care is actually relatively short compared to the delay of starting to worry about something and then seeking help.

Part of the problem with skin cancer is that you're looking at something on your skin and trying to figure out: has it changed? And then you're trying to work out whether it is serious enough, with no other symptoms, to go and speak to a clinician. And being very British — there was some research done where people were saying things like, oh, I didn't want to waste the doctor's time. If you've got a headache or pain, you're going to go solve that problem, but with skin the barrier to accessing care is quite high. And then when you layer on top the demand on our primary care services, it makes it even higher — because you have to be willing to wait two or three weeks to see a doctor. And if you're waiting two or three weeks, you know you're doing that because the demand on the doctor is high. So it loops you back into: have I got enough evidence to do this?

Ultimately, we think the real shift in skin cancer outcomes is going to come from being able to make it so frictionless that you can access care faster. That's our ambition. But we have to figure out this clinical part first, and make sure clinicians are comfortable in the role we play, because we're going to need their support. Going back to that earlier point — we want to make sure we have the pathway, what happens next, sorted out for the patient. So we need to do it with the healthcare system. That's a future, future goal of ours.

It's definitely a fascinating area. Even just on the image-based side, you could have it so your mirrors are constantly scanning and picking out these things, so you're not actively doing it —

Yeah — although it's still a little bit creepy. I still struggle sometimes thinking that my phone has all these pictures of me, because the camera's always on just in case I want to unlock it. So there are a few things to work through in terms of where the boundaries are between computers and humans. But there's a huge opportunity to make this more part of our life. The path to do that is relatively long, though, because it's just a really complex area. Diagnostics are complex in the sense that these last few percentage points of diagnostic accuracy come with a huge amount of work, and the more information we get from a high-quality image can be the difference between finding something versus not.

So we keep our eye on it. We're constantly thinking about where this is going. There are lots of liquid-biopsy ideas out there, and genetic biomarkers, which we think are great — but at the moment their performance isn't high enough to be competitive, and their cost isn't either, because you need a lab to develop or process a lot of these things. So the cost and the time is struggling to compete with an AI model where your cost is virtually nothing and your time is instantaneous. I think it's going to be a while yet before the wheel of innovation turns beyond image, from our point of view in skin cancer. But we're constantly thinking about it, because there are some really cool things that will happen in the future.

I just think there's so much scope to use technology like ours to redesign how we do healthcare — specifically, for us in the first instance, dermatology. If you started from scratch and said, okay, we've got these technologies now, how do we design dermatology services — where are the points where you really need humans involved and overseeing it to make sure you've got a great outcome? I think we want to be on the vanguard of that. I want us to be almost the digital primary care for dermatology: bringing a specialist function to be the front door of care around the world for all of dermatology, and to do that with our dermatologists, so we design it in a way that's hugely more efficient for them.

I'll give you another example. We obviously work in skin cancer, and we've talked a lot about the benefits of that. But if you look at something like acne, quite often patients come in and are laddered up different treatments until they get to the outcome. Laddering up treatments is difficult when you can only see a patient every six weeks or so — and that's quite a regular visit in a dermatologist's current workload. But not all of that needs to be. You could automate a lot of it. Instead of waiting six weeks and taking a point-in-time snapshot — where were they six weeks ago, where are they now, relying on your memory or an image you've taken — you could be collecting that continuously and saying, okay, we know this patient's responding to treatment, great, we don't need to see them again unless we're not getting to the outcome we expect. And much faster we can say, hang on, this patient isn't responding, we'll ladder them up — to the point where, for example, they might be on Roaccutane, which is quite intensive management from a patient point of view, making sure there aren't unintended pregnancies or that their blood work doesn't go strange.

So there's a huge opportunity in a number of different areas — in a different way to how we solve it with skin cancer — to really drive efficiency for a dermatologist and get them to the point where they can deliver the care they really want to deliver, and solve patients' problems sooner. Given that dermatology is such a massive part of clinical care — one in five appointments involves something to do with the skin — it feels like a huge opportunity to really use technology to show healthcare how you could do it differently.

No. I'm a real believer that we tend to get a sample size of one and try to draw information out of it — oh, people do this and then they do well because they've done this. And I just don't think that's true. The one thing I see in every entrepreneur I've worked with is resilience: the ability to keep getting back up when you've been knocked down, and you need to do that a million times to grow a business. But apart from that, there hasn't been any other thing I've seen. People say, I get up at 5am and I work out and then I do my emails and I make sure I don't have an inbox with more than five unread emails — and I've heard a million things like that, and I don't see any of them making a difference. I think it's just: be yourself, be resilient, do what you think is right, and try your best to achieve what you're setting out to achieve. That's it really.

Yeah, there are lots of things. I don't think we want to dive into the insecurities of a founder — there are a million of them, we feel like we're imposters half the time. But one thing that stands out is that I'm a big-picture person. I get excited by what we can do, I get excited talking with people about all this future we're going to create — and then I find it very difficult to get down into the detail of, okay, to deliver this we're going to need to do these five things in this order, and phase it this way, and staff it this way. So I've been surprised I've been able to get this far just with the energy to push for this new future we're trying to build. I've been really lucky that we've got some great people in the company who can pick up the pieces and say, okay, this is how we're going to make this happen. I think many entrepreneurs probably have this trait. But I'm still surprised we've got so far — given how much our business is built on detail — with someone like me who's not great at it.