AI is already outperforming humans in skin cancer diagnosis.
Image: Supplied
Caitlin Carlson
A future where artificial intelligence will guide skin cancer detection is in sight, experts say, even as human care remains essential.
“The whole ecosystem has really matured and we are now past the hype and the big media headlines,” says Ivy Lee, a Los Angeles-based dermatologist who chairs the Augmented Intelligence Committee at the American Academy of Dermatology.
A 2017 paper in the journal Nature found that an AI model that reviewed 129 450 clinical images outperformed 21 dermatologists when diagnosing skin cancer.
While promising, the paper didn’t consider that a human dermatologist will ask the patient questions about the lesions and touch and feel them, says Veronica Rotemberg, a dermatologist at Memorial Sloan Kettering Cancer Center in New York. “There’s a lot more about clinical medicine than looking at a photo.”
Fast-forward to today, and experts are determining how these AI models can assist physicians and help patients in real life. “We’re questioning how we can use it practically and meaningfully in the real world,” Lee says.
“Overall, I’m optimistic and hopeful,” she adds. “We know we need to do better for more people.”
Handheld devices, which a health-care provider hovers over lesions for analysis via optical spectroscopy, can assess for the three most common skin cancers - basal cell carcinoma, squamous cell carcinoma and melanoma. This may help cut back on unnecessary dermatologist referrals and biopsies in underserved areas, says Mitchell A. Kline, a clinical assistant professor at New York-Presbyterian Weill Cornell Medical Center.
The devices are intended for general practitioners and other health-care professionals who don’t specialize in dermatology, but Kline is using them anyway to occasionally assist with a tough call over whether to biopsy something and for data collection. He says the real potential value of the tools is for clinicians whose ability to diagnose is less reliable and who don’t have access to other diagnostic devices.
One such product, DermaSensor, funded research that was published in the Journal of Primary Care & Community Health and pointed to promising early data on its use by primary care physicians.
DermaSensor did a reasonable job of picking up skin cancers, but it also had a high rate of false positives and does not work with all types of lesions, according to Roxana Daneshjou, an assistant professor of biomedical science and dermatology at Stanford University. But her biggest concern, one echoed by other experts, is whether such models have been tested or trained on diverse skin tones.
Other companies in the marketplace include Skin Analytics, approved in the European Union as a medical device, and Nevisense, another U.S.-based company whose device gained Food and Drug Administration premarket approval for detecting melanoma.
Sloan Kettering and other institutions use 3D scanners to photograph and monitor patients with complex melanoma risk.
“The next time they come in, you’re monitoring all those lesions for change,” Rotemberg says. “Patients have access to the images when they’re performing their self-examinations, so they can also either reassure themselves or flag suspicious lesions for their doctor to take a closer look at.”
Today, AI is used as part of the 3D body scan for rendering the images and image reconstruction, Rotemberg says. The 2.0 version of this would entail bringing in AI to automatically detect skin cancer, she says.
“We need a lot more research, but we’re actively working on it,” Rotemberg says. “We recently published a dataset of over 400,000 tiles cropped from total-body photography so that people can try to develop algorithms,” she says. Because of the volume and quality of the photographs, they’re “super challenging” for a dermatologist to interpret, she explains. “So this is a huge opportunity for AI.”
Some companies offer services like this already, but the experts note that they have not seen any publicly available data that proves their efficacy or regulatory status in the United States or abroad. Daneshjou adds that there may be companies doing trials that would then apply for FDA clearance. “I do think it is coming down the pipe.”
Lee also says there could be a future where these full-body scans are made widely accessible - and that would be a positive. “Anything to improve access to screening and to high-quality care is something we need to do better at,” she says. That would require solid scientific evidence that shows it’s safe, cost-effective and equitable, as well as regulatory work to ensure it’s sustainable and scalable.
But experts advise against using apps and chat bots to diagnose skin cancer.
“It’s actually quite problematic because [these apps] have not cleared any regulatory hurdles, and some of them are making claims around being able to diagnose skin cancer and there’s no actual available data to back up their claims,” says Daneshjou, who co-wrote a paper on the topic in JAMA Dermatology in 2024 with Rotemberg and others.
Mobile apps and chatbots may be “dangerous because oftentimes we turn to these tools as patients because we can’t get access to health care,” Lee says, adding that data privacy and security should also be a concern.
Experts also do not recommend using ChatGPT and other large language models to analyze your skin lesions. “What we have found from prior work is that these models don't do that great on skin cancer diagnosis,” Daneshjou says.
Rotemberg says direct-to-consumer products remain the “furthest away” from being useful for patients, although it’s “a very promising area.”
As of now the gold standard is to get an annual skin check with a dermatologist, Lee says. And if you notice any new or changing spots, you should see your dermatologist or physician. | The Washington Post