Julian Abhari (CTO), Dani Marques (CEO), and the SkinCheck team are on a mission to make skin cancer screening more equitable — and we’re honored to support them as part of the Arionkoder Reshape Health Grants.
Their AI-powered app helps detect skin lesions across all skin tones, tackling the biases that often go unaddressed in dermatology. One of the key areas we’ll be working on is helping the model perform better in varied real-world settings, like poor lighting or irregular image capture.
Proud to partner with innovators making early detection more accessible to all.
See the transcript below:
Hello. I’m Julian Abhari. I’m the cofounder and CTO at Skin Check.
I actually started building SkinCheck, after my mom battled multiple cases of skin cancer growing up during my childhood. And it was in high school that initially, I was really interested in machine learning and and app development, and I created a computer vision prototype of an algorithm that would be able to detect skin cancer patterns.
Currently, we face on our AI technology about three crucial issues that require far more research and development. One is in regards of stabilizing computer vision variability, basically enabling us and our AI technology to standardize across the position, the angle, the distance, the lighting. And this will be built out entirely in the scanning functionality of our platform, which requires a lot of UX and UI research and development. And then at the same time, will require us to ensure that our models are extremely robust to adversarial attacks and, of course, any variability in the position, angle, distance, lighting, to ensure that our AI is as robust as possible. That’s one particular challenge that’s really important for us to go forward.
Another one is specifically in maximizing the sensitivity and specificity scores for skin cancer overall, which requires us to improve and expand our classifications to multiple different types of skin lesions. And then finally, enabling us to mitigate racial biases, it’s a really core challenge that we’re able to effectively create and maximize our usage of the small minority of datasets that include darker skin tones for us to ensure that representation from communities of Pacific Islander, African, African American are sufficiently represented and, that our models are able to check and correct the learned differences.