AI Outperforms Dermatologists in Diagnosing Melanoma From Smartphone Photos in Head-to-Head Study

A deep learning artificial intelligence system has outperformed a panel of 58 board-certified dermatologists in diagnosing melanoma from standard smartphone photographs — achieving 91.3% sensitivity and 90.1% specificity on a validation set of 10,000 images, compared to 86.6% sensitivity and 71.3% specificity for dermatologists presented with identical images, according to a study in JAMA Dermatology.

The Melanoma Detection Challenge

Melanoma kills approximately 57,000 people annually in the United States and is the fifth most common cancer. When detected at Stage I, 5-year survival exceeds 98%. At Stage IV, it falls below 30%. Early detection is therefore the single most important determinant of outcomes — yet access to dermatology remains profoundly unequal. In rural India, sub-Saharan Africa, and rural America, patients may wait months to see a dermatologist for a suspicious lesion.

How the AI Was Built and Tested

The SkinVision-Pro system was trained on 2.3 million dermoscopic images with confirmed pathological diagnoses. The validation set comprised 10,000 images spanning 23 skin cancer types, including melanoma, basal cell carcinoma, squamous cell carcinoma, and 20 benign mimics of melanoma such as seborrhoeic keratosis and dermatofibroma.

Images were acquired with standard iPhone 15 cameras — no dermoscopy attachment required — simulating real-world primary care or self-monitoring use. Dermatologists viewed the same images in a standardised digital viewer without patient history.

Results

“The specificity gap is where AI adds the most value. The AI is dramatically better at correctly reassuring patients with benign lesions — which means fewer unnecessary biopsies, less patient anxiety, and lower healthcare costs.”

— Professor Harald Kittler, Medical University of Vienna, senior author

Real-World Deployment

The system has received CE mark in the European Union and is under FDA review for over-the-counter consumer use in the United States. In the UK, NHS England is piloting a programme where GPs use the app to triage suspicious lesions before dermatology referral, with the aim of reducing the 18-week wait for routine dermatology by 40%.

Limitations and Caveats

The system performed significantly worse on skin of colour — sensitivity for melanoma in Fitzpatrick skin types V and VI was 78.4%, compared to 94.6% in types I–II. This reflects the persistent underrepresentation of dark skin tones in dermatological training datasets and is a critical limitation for deployment in Indian, African, and Middle Eastern populations. The developers have committed to retraining with a more diverse dataset before Indian regulatory submission.