The problem with using AI to detect skin cancer

Artificial intelligence is trained to detect skin cancer using photos — but they're mostly photos of white people.

Maija Kappler 4 minute read November 12, 2021
skin cancer AI

Artificial intelligence has a lot of potential to detect skin cancer — but there are also some significant drawbacks. (Getty)

At a time of limited access to burned-out doctors, mass exodus of nurses, increasingly long patient waitlists, and an ongoing pandemic, the use of artificial intelligence in the medical system has massive potential. Being able to diagnose — or at least flag — significant issues before setting foot in a doctor’s office would help thousands of people. And the possibility of AI diagnosing something with visible physical symptoms would be a natural place to start.

But unfortunately, with skin cancer, the existing AI just isn’t good enough.

That’s what researchers from the U.K.’s National Cancer Research Institute (NCRI) discovered when they examined the way AI learns about skin cancer, in a study published this week in The Lancet.

“AI programs hold a lot of potential for diagnosing skin cancer because it can look at pictures and quickly and cost-effectively evaluate any worrying spots on the skin.” Dr. David Wen of University of Oxford told the NCRI’s news outlet. “However, it’s important to know about the images and patients used to develop programs, as these influence which groups of people the programs will be most effective for in real-life settings.”

AI learns what it’s looking for through “training”: being fed thousands of images of skin lesions and other abnormalities that have already been determined as either cancerous or non-cancerous. In theory, the program can then track and flag the similarities that may lead to a cancer diagnoses in new photos.

Researchers found and examined 21 sets of training material data that the AI had received, which together included more than 100,000 pictures. But they realized a lot was missing in how the AI was trained: few photos were taken in the highly specific way that would be most helpful (one photo of the potentially cancerous lesion, along with a photo taken with a special hand-held magnifier). And there was no information provided about how or why the images were chosen to be included.

And, more significantly, they didn’t have sufficient racial data for the program to recognize what skin abnormalities might look like on a diverse range of skin tones.

Most of the photos — 14 of the 21 data sets — provided information about the country the photos came from, but few of them were specific about the patients’ skin colour or ethnicity. And the images that did specify skin colour were overwhelmingly white: out of 2,436 photos with that information, only 10 were of people with brown skin, and there was only one single photo of dark brown or black skin. Even fewer photos (1,585) — specified ethnicity. There were no photos of people whose ethnicities were African, Afro-Caribbean or South Asian.

When it comes to something as specific as skin cancer, that’s a significant problem, Wen explained.

“Research has shown that programs trained on images taken from people with lighter skin types only might not be as accurate for people with darker skin, and vice versa,” he said. “This can potentially lead to the exclusion or even harm of these groups from AI technologies. Although skin cancer is rarer in people with darker skins, there is evidence that those who do develop it may have worse disease or be more likely to die of the disease. One factor contributing to this could be the result of skin cancer being diagnosed too late.”

Dr. Neil Steven, who was not involved in the research, but who is a member of the NCRI Skin Group, said it’s worrying that many non-white people have essentially been left out of the AI skin cancer training so far.

“The findings of this review… raise concerns about the ability of AI to assist in skin cancer diagnosis, especially in a global context,” he said. “We already know that there are not enough pictures of people from black and Asian backgrounds in the textbooks we use to train doctors… I hope this work will continue and help ensure that the progress we make in using AI in medicine will benefit all patients, recognizing that human skin colour is highly diverse.

Wen added that he and his colleagues plan to improve the standards used in training and developing AI for medical use, with an emphasis on the importance of including a wide variety of skin colours and ethnic backgrounds in the patients whose images are used.