Deep Ensemble for the Recognition of Malignancy, known as DERM for short, is an artificial intelligence (AI) that its creators claim is as accurate as a dermatologist in diagnosing skin cancers. Created by Cambridge-based life sciences company Skin Analytics, DERM aims to enhance the decision-making of GPs when it comes to diagnosing potential skin cancers in primary care.
Skin Analytics has created a smartphone-compatible device that takes a dermascopic image of a skin lesion, which is then fed into DERM’s deep learning algorithm. Just as human beings learn from experience, DERM’s deep learning algorithm trains itself to diagnose cancers by performing its analytical task repeatedly and improving in accuracy each time.
Skin cancer is one of the most common cancers in the world with almost 16,000 cases diagnosed in Britain every year, 86% of which are preventable.
Chloe Kent: How was Skin Analytics founded?
Neil Daly: I started Skin Analytics in 2012. I grew up in Australia, where there are huge numbers of people diagnosed with skin cancer and a lifetime risk of two thirds, so we’re taught about it from a very early age. I had a background in physics and maths, and I found the nascent market of artificial intelligence fascinating. It seemed to me that there was a really interesting opportunity to try and start applying artificial intelligence to visual diagnostics.
Skin cancer seemed like a really good place to start, so I went around and I asked every dermatologist and doctor and researcher who would speak to me if the idea was any good. Most of them said no, but eventually a professor of software engineering up in Cambridge by the name of Bill Fitzgerald thought that it could be a really good idea so that’s how we got started.
CK: How does Skin Analytics’ platform work?
ND: It’s a deep learning AI. We actually started out using classical AI and we did that because we wanted to measure the same factors a dermatologist would, using pre-existing codified datasets. We did this for two or three years and it worked reasonably well, but what we found out very quickly is there’s a lot of learning that dermatologists do about skin lesions which they find hard to codify themselves and that’s where deep learning works so well. So we’ve built our own network from the ground up specifically to deal with skin lesions.
We get a panel of consultant dermatologists within the NHS to report for us, and what we’re going to start doing now is running AI in parallel showing the results then phase out the dermatologists and phase in the artificial intelligence.
CK: What would be the experience of a patient when they have a mole run through Skin Analytics’ system?
ND: We finished up our perspective clinical study in hospitals at the end of last year, so at this point we’ve registered as a class of medical device and we’re only now starting to roll out the artificial intelligence side of things.
Effectively, patients will be able to visit a GP who has been given our equipment and the GP will take a dermoscopic image of a concerning mole by attaching a lens to a smartphone using a magnetic clip. The image of the skin lesion will then be uploaded to our servers in the cloud and within about ten seconds a result will come back to the clinician which explains what DERM thinks the image shows as well as the performance characteristics of the AI, how sensitive and specific the technology is for the clinician to then be able to make a better decision.
CK: What are the benefits of this AI method versus more commonly used techniques of skin cancer diagnosis?
ND: If you’re looking specifically in the UK, there are two distinct problems here.
Firstly, people are seeking medical attention later than they could be, so the lesion has advanced further and survival rates drop significantly as the cancer progresses. The second key issue is that when people do go to their GPs, the doctors themselves aren’t as good as a trained dermatologist at spotting skin cancers. So what they tend to do is miss cancers in some patients and over-refer a significant number of others.
Within the NHS there are two different referral pathways that can be used. One of them is if it’s not suspected to be a melanoma, as the non-melamona skin cancers very rarely result in death. The other one is a two-week wait pathway, which means the patient has to be seen in the hospital within two weeks. That’s a politically measured timescale for all cancers. It creates a huge amount of pressure on the NHS, costs a huge amount to meet and results in a whole bunch of expensive initiatives to reduce that two-week wait time.
GPs tend to use the two-week wait referral for a lot of cases, significantly more than they should. Anecdotally, some hospitals we work with would tell us that 90% of the two-week wait referrals that they see shouldn’t have been in the hospital at all. So what we’re trying to do with our technology is get a dermatologist-quality system rolled out into primary care, one where you won’t miss a melanoma and can identify what a lesion is much more effectively.
This way we can help the GP figure out what the right pathway. The vast majority would be non-urgent refers, which would take a lot of pressure off the healthcare system.
CK: How have Skin Analytics’ services been used in the field so far?
ND: We have four partners at the moment. Our first partner was a private insurance company called Vitality Health, and we work with a group of private GP practices across London called DocTap, which has 11 surgeries across the city. Within the NHS we have two deployments, one in West Hampshire CCG [Clinical Commissioning Group] at a community clinic, and one in GP surgeries across Norwich.
CK: What would you say is your greatest achievement is so far as a company?
ND: I think that we made a decision very early on that we were a medical company not a technology company. What that has meant is that we’ve spent a lot of our time, effort, and funds on going out and capturing clinical evidence in a really rigorous and difficult way.
It almost broke us a number of times and convincing the NHS to do a study for us was particularly difficult, but at the end of the day what we’ve got is some really strong evidence that Skin Analytics’ platform works. To take technology like ours, artificial intelligence, and embed it into the healthcare system, you really need that evidence that we’ve been building and I think that’s probably the best thing that we’ve done.
CK: What does the future look like for Skin Analytics?
ND: The future’s quite bright for us at the moment. Our clinical study is due for publication shortly, and we’re in conversation with a number of healthcare providers in the UK. We’re also getting a lot of inbound requests from places like the US, Australia and Germany for the technology. The whole area has really exploded for us in the last six months.
CK: What would you say is the most important thing for people to be aware of in being mindful of skin cancers?
ND: By far the most important thing for people to be considering, especially leading into summer, is if you see something that looks like it’s changed or you’re uncomfortable in any way then you need to go and speak to a doctor about it. We know from the literature that people are resistant, they don’t want to waste their doctor’s time; there are very few symptoms of an early-stage skin cancer so it’s really hard to get people to do that but it’s absolutely what needs to happen.