The AI ​​model predicts skin cancer risk better than current methods

ENGINEERINGNET.BE – 'As a dermatologist, I care about things like pigment spots, wrinkles or the fact that someone lived in a sunny country. “The algorithm captures more characteristics than just traditional skin cancer risk factors,” says dermatologist Dr. Marlys Waki.

The Erasmus MC team trained the AI ​​model with more than 2,800 facial images, each taken under the same conditions. Researchers know of people who developed skin cancer after the photo was taken.

The algorithm extracted 200 features from the images, based on which the risk of skin cancer, on the face or elsewhere on the body, was predicted. AI prediction is better than prediction based on skin research, questionnaires and genetic research.

The researchers used so-called explainable AI methods. This means that this can be determined based on what the algorithm bases its predictions on. This is not always the case with artificial intelligence. It is often a black box where it is not clear how the model arrives at a particular result.

“We make visible the 200 characteristics on which the model bases its predictions. If we try to translate that back to what is known about skin cancer risk, we see that the algorithm extracts features from faces that we already know are associated with skin cancer. Think about wrinkles, for example Example,” Waki ​​explains.

But the model also captures features that are less obvious to people. “Explainable is not the same as explainable.” “We can explain what the algorithm does, but its meaning is not entirely clear,” said co-researcher Dr. Gennady Roshchebkin.

Wakkee and Roshchupkin see two potential roles for the AI-based skin cancer risk model: prevention and awareness. “If you know from the photo that you have a high risk of skin cancer, you will use the sun more consciously. We also hope that people who are wary will watch their skin more closely.”

Currently, the AI ​​model cannot yet be used in the clinic, but the focus is on follow-up research. “For example, we want to test the algorithm on larger datasets containing images of people of older ages and ethnic backgrounds. By sharing the model publicly in the research community, fellow scientists can build on it. This could accelerate the AI ​​community,” Roshchobkin said. entire”.