American Melanoma Institute · 5 min read
What artificial intelligence devices and phone apps can and can't do for skin cancer detection today, based on FDA decisions and published studies.
Computer programs can now sort images of skin spots with impressive skill in research settings. That progress is promising, and it has also produced a crowded market of phone apps with big claims. The products fall into 2 very different kinds, and the evidence behind them differs just as much.
The first kind is a medical device used by a doctor during a visit. These go through review by the US Food and Drug Administration (FDA), and the FDA spells out who may use them and on which patients.
The best-known recent example is DermaSensor, which the FDA authorized in January 2024 through its De Novo pathway for new types of devices. The device is a handheld probe that shines pulses of light into a spot and analyzes how the light scatters off the cells, a technique called elastic scattering spectroscopy. An algorithm then returns a result that either flags the spot for further investigation or does not.
The FDA authorization is narrow. The device is for physicians who are not dermatologists, mainly primary care doctors deciding whether to refer a patient. The authorization covers patients aged 40 and older, for spots the doctor already finds concerning. The FDA's summary states that it should not be used as the sole basis for a diagnosis or to confirm one.
The device is also not meant for certain spots. Excluded are lesions on the palms, soles, genitals, or mucosal surfaces, under the nails, near the eye, within tattoos, and lesions that were biopsied before.
In the main study supporting authorization, primary care doctors at 22 centers tested 1,579 lesions on 1,005 patients, and every lesion was biopsied. The device flagged 95.5% of the skin cancers. That sensitivity was higher than the doctors' own assessments in the same study.
The trade-off was specificity, which was 20.7%. In plain terms, the device also flagged about 4 out of 5 harmless spots. A negative result was reassuring, since 96.6% of spots the device did not flag were benign. A positive result by itself said little.
An analysis in npj Digital Medicine raised a second issue: 97.1% of the study participants were White. The FDA required the company to run further testing after authorization in groups that were underrepresented, and its summary advises relying mainly on clinical judgment for patients with darker skin types until more data exist.
The second kind of product is a consumer app that analyzes a photo you take at home. The most thorough review of these is by Freeman and colleagues (BMJ, 2020), who searched for every study testing an algorithm-based app against a confirmed diagnosis.
They found 9 studies covering 6 apps. The studies were small and had serious design problems. In most, a clinician chose the spots and took the photos, which doesn't reflect how people use an app at home. Many images couldn't be analyzed at all and were dropped from the results.
Of the 2 apps that carried a European CE safety mark at the time, one missed all 5 melanomas in its only study. The other identified 80% of cancerous or precancerous spots and correctly cleared 78% of harmless ones. The authors concluded that current apps can't be relied on to detect all melanomas, and that performance in everyday use is likely to be worse than the studies suggest.
In the United States, the American Academy of Dermatology points out that many diagnostic apps avoid FDA review by labeling themselves as informational. A developer who does that has no obligation to prove the app is accurate.
An algorithm trained on sharp, well-lit, professionally framed images meets something different in a bathroom: blur, shadows, hair, and odd angles. It sees one spot in isolation. A dermatologist compares that spot with every other mole on your body, feels for firmness or tenderness, and asks how it has changed.
Some melanomas are pink or skin-colored with little pigment, and these are hard for people and for software. Training data also matters. If a system learned mostly from lighter skin, its accuracy on darker skin is unknown until someone tests it.
The biggest risk is false reassurance. A person who gets a low-risk reading on a changing mole may put off a visit, and with melanoma, months can matter.
The research gives good reasons for optimism. Image-recognition systems keep improving, and researchers are testing them the right way, in prospective studies on the patients and settings where they would be used. Several groups are studying AI as a second reader alongside a clinician, and early work suggests the pairing can do better than either alone.
Wider and more diverse image collections are being built to address the skin-tone gap. More data is needed before anyone can say how these tools change outcomes such as melanoma thickness at diagnosis or the number of unnecessary biopsies.
For now, the path to a diagnosis hasn't changed. A clinician examines the spot, usually with a dermatoscope, and a biopsy read by a pathologist gives the answer. No app or device on the market replaces either step.
