AI and skin cancer, honestly assessed

What the FDA has actually authorized

On January 12, 2024, the FDA authorized DermaSensor through the De Novo pathway, the first AI-enabled skin cancer detection device cleared for use in US primary care. The decision created an entirely new device classification: 21 CFR 878.1830, product code QZS, Class II, described as a “software-aided adjunctive diagnostic device for use by physicians on lesions suspicious for skin cancer.”

Read that description closely, because every word of it is doing work:

• “For use by physicians.” It is not a consumer product and was never authorized as one.
• “Adjunctive.” It supports a clinician’s judgment. It does not replace it and is not authorized to.
• “On lesions suspicious for skin cancer.” A doctor has already decided the lesion warrants attention. The device does not do the finding.

The evidence required to get there is worth stating too, because it is the thing consumer apps skip. DermaSensor’s pivotal trial ran across 22 primary care centers with 1,005 patients and 1,579 lesions; a second study added 311 patients and 440 lesions; a clinical utility study covered over 10,000 lesions and 108 physicians. The classification’s special controls require demonstrated accuracy superior to unaided users and a minimum 90% sensitivity for lesions with high metastatic potential, plus a postmarket study in underrepresented populations. The company behind it was founded in 2009 and raised roughly $43 million to get there.

How consumer apps actually performed when tested

Dermatologists at Memorial Sloan Kettering evaluated commercially available melanoma-identification apps. The summary from Dr. Veronica Rotemberg: “On average, they don’t do very well.” The apps tested reached roughly 59% accuracy.

A separate study in The Lancet Digital Health found that algorithms performed dramatically worse on image types they had not been trained on, around 6% accuracy on unfamiliar categories. Stray hairs and pen marks in an image measurably degraded results.

This category has already been prosecuted once

On February 23, 2015, the Federal Trade Commission settled actions against the marketers of two melanoma apps, MelApp and Mole Detective. Users photographed a mole, entered some information, and the apps returned a melanoma risk rating of low, medium, or high. The FTC found the claims lacked scientific support.

The orders bar the defendants from claiming that any device detects, diagnoses or assesses the risk of melanoma without “competent and reliable scientific evidence in the form of human clinical testing.” That phrase, assesses the risk of, is the one to remember. It means a risk score is treated exactly like a diagnosis. An app that carefully avoids saying “melanoma” while returning a red, yellow, or green rating is making the same claim in different clothes.

Where the FDA draws its line

Congress carved out a category of clinical decision support software that is not regulated as a device. Skin lesion analysis does not fit inside it, and the reason is the very first condition.

The FDA revised its Clinical Decision Support Software guidance on January 6, 2026, superseding the 2022 version. The revision loosened several things. It now allows a single recommendation rather than a list of options, for instance. But Criterion 1 was left standing: to qualify for the exclusion, the software must not “acquire, process, or analyze medical images, IVD signals, or patterns from signal-acquisition systems.” Analyzing an image is disqualifying on its own, whatever else the software does.

The agency’s Digital Health Policy Navigator says the same thing from the other direction, and expressly counts dermatology images as medical images. There is no framing that makes a mole-reading app not a medical device. There is only doing the clearance work or not doing it.

Where AI genuinely helps today

None of the above means the technology is useless. It means the useful versions sit in clinical hands:

• Adjunctive assessment in the clinic. Cleared devices used by a physician on a lesion already judged suspicious, where the measured benefit is accuracy above the unaided clinician.
• Serial comparison in total body photography. Software comparing today’s images against a baseline to flag what has changed. This plays to what algorithms are actually good at, which is difference detection rather than diagnosis, and the reported effect is fewer biopsies with a higher yield.
• Triage in settings with no dermatologist. An active research area, and the place the technology may eventually matter most.