The AI Dermatologist That Diagnoses Melanoma From a Smartphone Photo Is Now More Accurate Than Board-Certified Specialists

Dermatologists discuss a study that was published in Nature with a mixture of professional deference and something more difficult to identify. Nearly 130,000 clinical photos of skin lesions were used to train a deep learning algorithm. The algorithm outperformed a panel of 21 board-certified dermatologists when asked to categorize images as benign or melanoma. Not a little bit better. significantly improved, as determined by sensitivity and specificity over thousands of diagnostic choices. 2017 saw the publication of the study. The systems performing this work are significantly more capable now, almost ten years later, and the discussion in clinical medicine has subtly changed from “can AI do this?” to “how do we deploy it responsibly?”

In the list of human diseases, melanoma is an uncommon type of cancer. For the most part, it is a visible cancer that sits on the skin’s surface and may be detected by a keen eye before it has caused significant harm. Its mortality figures are particularly concerning because of this visibility. Although melanoma only makes up a small portion of all skin cancer cases, it is the cause of about three-quarters of skin cancer deaths. The timing of detection is nearly the only factor contributing to the disparity between incidence and mortality. Melanoma has a five-year survival rate of about 99 percent if detected early. That rate drastically decreases once it has spread to distant organs. In the most practical sense, AI is providing a way to move more patients into the earlier column.

The system currently furthest along in clinical deployment is DERM — Deep Ensemble for Recognition of Malignancy — which England’s National Health Service began piloting across several hospitals in 2025 as part of a three-year evaluation. DERM analyzes skin lesion images and generates a prioritization recommendation: which patients need to be seen first, which can wait, which lesions require urgent dermoscopy and biopsy. Early research showed that the system had a sensitivity of more than 90% for melanoma. According to extensive audit studies, board-certified dermatologists typically score between 82 and 85 percent on the same metric. The gap isn’t enormous, but in a system processing millions of referrals a year, it adds up to a substantial number of missed early-stage lesions.

The stakes are particularly tangible in the context of the NHS. Referrals for suspected melanoma in England have roughly doubled over the past decade, from around 450,000 in 2013 to more than one million in 2023. The capacity of specialists has not kept up. In some regions, patients are waiting over three months for an initial dermatology review — a gap that is not clinically neutral when the lesion in question is one that rewards early action. AI triage doesn’t solve the capacity problem, but it changes the nature of it: instead of a three-month queue where cases are seen roughly in the order they were referred, an AI screening layer can sort the queue by urgency, ensuring that the cases most likely to be serious move faster.

AI Melanoma Detection — Key Facts & Reference

The FDA authorized DermaSensor in 2024 — the first AI-powered dermatologic device cleared specifically for use by primary care physicians rather than specialists. That regulatory step is significant, not just administratively but in terms of what it signals about where the technology is heading. The original clinical model for melanoma detection requires the patient to have enough concern about a lesion to seek care, the primary care physician to recognize that the lesion warrants referral, and then an available dermatologist to perform expert assessment. DermaSensor inserts a validated AI checkpoint at the second step in that chain, the point where most early melanomas are missed — by generalists who may examine dozens of moles in a year rather than thousands.

As this progresses, it seems as though the accuracy debate has largely been resolved and the more challenging issues are now related to implementation. In 18 of the 38 studies examined, AI outperformed dermatologists, and in another 12, it performed similarly, according to a meta-analysis published in Nature in 2024 that covered research conducted between 2013 and 2023. That technology has been proving itself for almost ten years and is currently navigating the considerably slower processes of regulatory approval, hospital integration, workflow redesign, and clinician trust-building. It is not a technology that is still proving itself.

The restrictions are genuine and worth mentioning. The DERM team is specifically addressing this issue in their NHS trial by testing across a wider range of skin tones and lesion types. AI systems trained primarily on images from lighter skin tones have demonstrated decreased accuracy on Fitzpatrick type V and VI skin. Image quality is crucial: a clear dermoscopy image taken with the appropriate attachment lens yields better results than a blurry smartphone photo taken in dim lighting. Additionally, uncommon variations of skin cancer that might not be adequately represented in training datasets continue to be a vulnerability. These are reasons to be careful about how and where the technology is used, not to discount it.

It’s difficult to ignore the fact that the nations advancing this the quickest aren’t doing so by substituting dermatologists. They accomplish this by providing a validated second opinion that can operate quickly and on a large scale to all other members of the healthcare system, including primary care physicians, nurse practitioners, and patients who have smartphones and legitimate concerns about a changing mole. The knowledge of the dermatologist is not being replaced. It’s being saved for situations that actually require it.