Leo Innovation Lab: Detect Skin Cancer

Challenge

Leo Innovation Lab wanted to develop an App giving users the opportunity to diagnose skin diseases based on a photo of skin. Being part of a healthcare company, it was imperative to Leo Innovation Lab that they properly estimated the uncertainty in the predictions in order to not mis-diagnose the user, but rather know when they needed to retake the photo or consult with a doctor.

Solution

Using Borch, our deep universal probabilistic programming language, we developed and tested Bayesian models that were both accurate and uncertainty-aware for the identification of skin diseases.

Borch-based models require much less data and extract more information per observation, contributing to better uncertainty estimation. This is a major improvement in the quality of the suggested diagnostic results compared to results based only on maximum likelihood, where overconfident predictions are abundant.

A simulated lesion photo. The model never sees the sliders, only the resulting pixels.

AsymmetryBorder irregularityColour variationDiameter 4.5 mmPhoto blur

Simulate non-skin photoReset

---

Predictive uncertainty—

—

Business Value

Accurately quantifying uncertainty in the diagnoses is imperative for predictions of images. It can be as simple as dealing with photos that don’t contain skin, finding out when a picture needs to be retaken from a different angle, or when a doctor needs to be consulted.