A LARGE international diagnostic study has found that while modern artificial intelligence (AI) models can outperform less experienced clinicians in skin cancer diagnosis, they remain unable to match the performance of expert dermatologists in realistic clinical settings.
Researchers evaluated the diagnostic accuracy of three AI systems alongside 652 physicians with varying levels of dermatological experience. The study used a dataset of 1,117 skin lesion cases reflecting everyday clinical practice, including rare and atypical presentations. Cases incorporated both clinical and dermoscopic images, as well as relevant patient metadata.
The AI systems assessed included a first-generation convolutional neural network (CNN) and two foundation models, PanDerm unimodal and multimodal. Human participants completed evaluations of 100 randomly selected cases from the same dataset.
Expert Dermatologists Outperform AI Models
Expert dermatologists with more than 10 years of experience achieved the highest diagnostic accuracy overall, correctly classifying lesions in 74.2% of cases. The PanDerm unimodal model followed closely with an accuracy of 72.2%, while the multimodal model achieved 66.3%. The CNN performed least well, reaching only 56.7% accuracy.
Importantly, the unimodal foundation model outperformed physicians with fewer than three years of dermatology experience, whose average diagnostic accuracy was 68.2%. However, the AI model was unable to surpass the performance of highly experienced specialists and achieved results comparable to dermatologists with three to ten years of experience.
Limitations of AI in Skin Cancer Detection
The study also highlighted the limitations of AI systems when applied to complex, real-world datasets. Although previous research has demonstrated strong AI performance under controlled conditions, the authors note that diagnostic accuracy often declines when algorithms encounter the diverse and challenging cases seen in routine practice.
Overall, the findings suggest that AI may be most valuable as a clinical support tool rather than a replacement for specialist expertise. The researchers propose that future dermatology workflows could benefit from human–AI collaboration, with AI assisting junior clinicians, supporting triage decisions, and helping reduce fatigue-related diagnostic errors. However, they conclude that expert clinical judgement remains essential for achieving the highest levels of diagnostic accuracy in skin cancer detection.
Reference
Anriot J et al. Limits of artificial intelligence models for skin cancer diagnosis in realistic settings. JAMA Dermatol. 2026; 10.1001/jamadermatol.2026.1492.
Featured image: sebra on Adobe Stock
- Author:
