India has around 11,000 dermatologists – less than one for its 100,000 people. A majority of the skin doctors are in metro cities (70%), according to the World Congress of Cosmetic Dermatology, skewing the distribution. With around 10-12% of the population estimated to suffer from skin conditions and dermatology an inherently visual speciality, it’s a market ripe for artificial intelligence-driven disruption.
That’s exactly what Cureskin, a deep learning startup in Bengaluru founded by two ex-Googlers, is taking a crack at. Its Android app (it’s not on iOS yet), which made it to Y Combinator’s summer 2017 batch, diagnoses six types of common skin conditions – pimples, acne, scars, dark spots, pigmentation, and dark circles – and recommends treatment regimens. We spoke to its founders, Guna Kakulapati and Ramakrishna R, to know about their vision for the one-year-old startup, understand why deep learning is a means to solve this problem, and what’s in their service pipeline.
Cureskin came about in October 2016, around the time when Kakulapati was exiting Bidstalk, an ad-tech startup that was acquired by AppLift, while Ramakrishna was closing his hyperlocal content sharing startup, IWe. Both have a background in AI-related work: Kakulapati spent six years at Amazon and five in Google, Ramakrishna close to 10 years at Google.
“In Google, we worked on spam detection and other machine learning (ML) aspects. Guna had also worked on ML in his previous adtech company, which was using ML-based bid optimisation,” says Ramakrishna, Cureskin’s CTO. “We wanted to do something in healthcare and we thought dermatology is a good place to start.” We are meeting at Starbucks in Koramangala, a southeast Bengaluru suburb with India’s highest concentration of startups.
In its first iteration, Cureskin started off as a pure consultation platform. “For AI, the main thing is data – most of the publicly available data sets are around cancer and skin cancer data. We tried our algorithms on them and were able to get dermatologist level accuracy,” says Ramakrishna. Skin cancer detection was not a suitable use case in India as its isn’t widespread here thanks to the protective effects of melanin in the population here.
The app has a pretty straightforward user experience: you take a picture of your face, the app scans the photo, and pinpoints skin issues such as dark spots, acne scars, or comedones. To treat the skin problems, the app’s chatbot asks a few questions and, depending on the inputs, offers an eight-week skincare regimen. It includes skin care products, bi-weekly reviews from Cureskin’s in-house dermatologist, real-time chat support, and diet and lifestyle guidance.
“The key thing is that we wanted to build something that is outcome driven. The key insight is that people, it’s not that they want to consult, or buy medicines. The real problem is that the user wants to get better, wants to cure him/herself,” says CEO Kakulapati. “What we’re building here is something that will give those outcomes, track those outcomes, and that is what we measure ourselves against.”
The odd thing about the app is that it uses the rear camera of the smartphone instead of the front-facing (aka selfie) camera. “It’s not only the resolution, but the focus, which is much better in the back camera,” says Kakulapati. The Cureskin AI needs to analyse the skin texture at a pore level of detail to figure out the skin type and skin sensitivity, along with skin issues. “Because people are going to make decisions based on the results and recommendations we make, we need very high-quality and in-focus images.”
What really is the deep learning under the Cureskin hood? “At the top level, we kind of do object detection: just like people detection, pedestrian and car detection. We treat each of these [skin conditions] as classes and objects, and do detection on that,” Ramakrishna says.
The startup claims that its deep learning algorithms are at dermatologist level accuracy for the six skin conditions it works on at present. Its goal is to be able to diagnose at least 60 skin conditions such as different types of dermatitis, pigmentation conditions like melasma, etc.
“We train neural nets (programs that mimic brain cells). Sometimes, for some class of problems, we can combine them, for some skin conditions we need to train them differently. For example, for bigger regions, when the neural network has to understand big patches of the face, then that’s a separate model,” Ramakrishna says. “If it [the neural network] has to understand smaller patches of face, for example comedones (which are tiny), that requires a different training model, compared to pigmentation, which are huge patches. So we try a combination of models, where one model is trained only for detecting smaller areas, and another model is detecting larger areas.”
While it’s easy to get 70% accuracy with deep learning, fine-tuning the performance to 80-90% accuracy requires knowledge, experience, intuition, and data, with the latter playing the most crucial part, says Ramakrishna.
I gave the app a try to get a hands-on experience of its monetisation funnel. It found dark spots on my face and then went on to recommend a skincare regimen package that cost Rs 2,160 that had three skincare products (facewash and soothing cream), dermatologist teleconsultation, an 8-week skin regimen plan, and a bi-weekly review. Ticket sizes start from Rs 1,800, the founders say, adding that there’s a money-back guarantee if users are not satisfied with the treatment.
Alphabet Chairman Eric Schmidt’s idea of a perfect startup is pretty close to the concept of Cureskin. He described an AI-based skin diagnosis tool that uses machine learning, and crowdsources inputs from actual dermatologists at a mechanical turk-like price ($1). Kakulapati contends that Cureskin is a different model from Schmidt’s vision, where he was talking about the acquisition cost of data. “Ours is not a B2B, it’s a B2C business model, so the data is collected directly from the users,” he says.
Cureskin has an in-house dermatologist and a couple of external dermatologists, who add a manual screening layer and provide users support and prescription medicines if needed. AI detection and chatbots help add scale to the number of patients they can handle, from 10-20 patients to 400 patients a day.
We reached out to startup data tracker Tracxn to find out if Cureskin has any comparable. Their analysts sent in a list of a dozen odd startups, none of which really seemed to be doing what Cureskin is. One reason for this might be that AI-first apps like Cureskin are hard to clone. “We work with deep learning techniques because that is the only machine learning algorithm that has been able to cut past human doctors in the medical field,” says Ramakrishna.
For now, Cureskin has a first-mover advantage in the space, with over 50,000 downloads on the Android app following its three-month stint at Y Combinator, where the founders were challenged to double their user base every week. It currently has 25,000 monthly active users, they say. An iOS app is coming as well but they don’t have a clear date for its launch yet. While the app has been launched for India, it could scale to other countries, as well.
“This is a virtuous cycle for us. As we do accurate detection, more people start using it, which gives us more data,” says Ramakrishna.
Earlier this year, Google said that it had trained a neural network to exceed the performance of a pathologist at detecting skin cancer.
“AI lets us identify both the class of the problem and the severity of the problem,” says Kakulapati. He lists three advantages: one, convenience because everything is on the phone; two, continuous care because the app continuously engages with you unlike a doctor who gives you attention only during a consultation; and cost-effectiveness because treatments can reach and scale up even where it is physically impossible.
Cureskin also sees itself as a platform, with the app functioning as a gateway. It could be launched as a Facebook Messenger bot, for example, or integrated with other platforms, such as skin clinics and hospitals, where a patient comes takes a photo in a computer kiosk, which produces a skin report.
Bengaluru dermatologist R Raghunath Reddy expressed his doubts as to whether a skin care app can be reliable. “I do have my reservations as a lot of skin conditions look the same,” he says, adding he’s not reviewed the app. A Chennai dermatologist felt if the condition is a simple one, the app will serve its purpose. “The downside is that it could lead to unnecessary panic and mistreatment if used inaccurately,” she said, asking to remain anonymous.
Other doctors saw potential in Cureskin. “One of the things about dermatology is that you can diagnose without too many laboratory interventions. That lends itself to AI quite easily,” says Aparna Santhanam, a dermatologist from Mumbai, who has worked on various apps and questionnaire-based dermatological diagnosis. “We used to be able to arrive at a probable diagnosis and treatment of your condition and put a notice at the bottom saying please consult a dermatologist,” she says.
When Cureskin says six types of skin conditions, “they’re probably covering 80% of the population, 80% of the time. The rest come under a smaller category,” Dr Santhanam says. Still, she adds, a virtual diagnosis will never be as accurate as feeling the skin physically through touch.
Charu Sharma, the director of dermatology at Cureskin, says while the in-house team handle skin queries that the AI cannot, they do come across cases where they recommend that a patient see a dermatologist when physical procedures are required.
Cureskin guides people to qualified care instead of self-medicating themselves. Topical steroid abuse is rampant in India. “Three of every ten patients in my consultations have been using, or continue to use without realising that one of the reasons for their damaged skin can be steroid abuse,” Sharma says.
Both Sharma and Santhanam say that topical steroids can be effective for a short time period. “It can’t be used for years and years. We [Cureskin] make sure that a good consultation and counseling is given prior to each regimen, follow up with the patient, to make sure that we are aware of when they’re using and when to stop it,” Sharma says.