AI-produced images can’t fix diversity issues in dermatology databases

Photo by: IMAGE POINT FR/NIH/NIAID/BSIP //Universal Images Group via Getty Images

Light skinned people are often biased towards image databases of skin conditions. One group wants to use artificial intelligence to fill in the gaps instead of waiting for the process of collecting more images to be completed. It is working on an artificial intelligence program to create synthetic images of diseases on darker skin and use them for a tool that could help diagnose skin cancer.

The ultimate solution is to have real images of darker skin.

Synthetic images could cause problems for other experts working in the field. Adding more diverse real images to existing databases is what the focus should be.

There are many attempts to use artificial intelligence in the field of skin care. Researchers use images of moles and rashes to figure out the most likely type of issue. The results can be used to make diagnoses. Most tools are built on images that don't include many examples of conditions on darker skin or have bad information about the range of skin tones they include. It's hard for groups to be sure that a tool will be accurate on darker skin.

That's why the team turned to synthetic images. The project has four main phases. The team analyzed available image sets to understand howunderrepresented darker skin tones were to begin with. Thanks to the advances in artificial intelligence and deep learning, we were able to use the available white scans.

The team will combine synthetic images of darker skin with real images of lighter skin to create a program that can detect skin cancer. It will continuously check image databases to find new, real pictures of skin conditions on darker skin that they can add to the future model.

The team isn't the first to create synthetic skin images, a group that included researchers from the health company published a paper in 2019. A tool that can identify skin conditions was announced last spring.

Synthetic images are a stopgap until more real pictures of darker skin are available. Even as a temporary solution, Daneshjou is worried about using synthetic images. The research teams would have to check if the images were normal and not strange. That type of quirk could skew the results. The only way to confirm that synthetic images work as well as real images in a model is to compare them with real images that are in short supply.

Daneshjou is concerned if a diagnostic model is based on synthetic images from one group and real images from another. The model could perform differently on different skin tones.

She says that synthetic data could make people less likely to push for real, diverse images.