Real-World Study Finds Experienced Dermatologists Outperform AI in Skin Cancer Diagnosis

Expert dermatologists with more than 10 years' experience outperformed 3 artificial intelligence (AI) diagnostic systems when tested against a realistic mix of common, rare, and atypical skin lesions, according to a diagnostic study published online June 3, 2026, in JAMA Dermatology.¹ Researchers led by Julien Anriot, MD, and Luc Thomas, MD, PhD, of Claude Bernard University Lyon 1 in Lyon, France, compared the performance of 3 AI diagnostic systems with that of 652 physicians using 1,117 standardized cases drawn from the Test of Dermoscopy for International Validation (TDIV) platform. Expert dermatologists achieved the highest multiclass accuracy, at 74.2%.¹

AI models have previously demonstrated diagnostic accuracy comparable to or exceeding that of dermatologists under controlled research conditions. However, their performance against a broader mix of common, rare, and atypical skin lesions has remained less well characterized. To address that gap, investigators compared physicians with varying levels of dermoscopy experience against a first-generation convolutional neural network (CNN) and both unimodal and multimodal versions of the foundation model PanDerm.

The TDIV dataset paired clinical and dermoscopic images with patient history, demographics, and risk factors across 1,117 cases, deliberately including rare and atypical tumors known to challenge clinicians. Physicians completed 1,092 diagnostic test iterations, and the primary outcome measured multiclass accuracy across 9 diagnostic categories.¹

Experienced Dermatologists Achieved the Highest Multiclass Accuracy

Experts remained the highest-performing group, achieving 74.2% multiclass accuracy.¹ The unimodal, image-only configuration of PanDerm reached 72.2% accuracy, statistically comparable to dermatologists with 3 to 10 years' experience and higher than readers with less than 1 year of experience, who scored 59.1%.¹ The CNN trailed every physician group at 56.7% accuracy, the lowest of any system evaluated. Overall, the 652 participating physicians achieved a mean multiclass accuracy of 65.9%, below the top-performing AI configuration but above both the CNN and the multimodal PanDerm model.³

AI Excelled in Binary Classification but Showed Limitations

For the simpler task of distinguishing benign from malignant lesions, the unimodal PanDerm model achieved the highest balanced accuracy, at 0.82, compared with 0.65 for physicians overall.¹ This advantage was driven largely by specificity: the unimodal model correctly classified benign lesions in 94% of cases, while the multimodal version, which incorporated clinical photographs and patient metadata, reached 97% specificity. Compared with AI, physicians demonstrated higher sensitivity but lower specificity, reflecting a greater emphasis on avoiding missed malignancies. The most experienced dermatologists maintained the highest sensitivity of any group.¹

Adding clinical context reduced rather than improved PanDerm's multiclass performance, with the multimodal configuration achieving 66.3% accuracy compared with 72.2% for the unimodal version.¹² The authors attributed part of this decline to a distribution shift between the model's training images and the more complex clinical photographs included in the TDIV dataset, along with an apparent underrepresentation of acral melanoma among the malignant lesions both configurations failed to identify. They also noted the reader population was predominantly French, the patient cohort was largely of European ancestry, and darker skin phototypes were underrepresented, limiting the study's generalizability.¹

Authors Highlight AI's Role as Clinical Decision Support

The investigators emphasized AI's role as a clinical support tool rather than a replacement for physician expertise. Anriot and Thomas wrote, "AI systems demonstrate strong potential as diagnostic support tools, particularly for early-career clinicians."¹

The authors suggested human-AI collaboration may represent the most practical path for clinical implementation, with AI serving as a safety net and educational resource for novice clinicians while functioning as a systematic second reader for experienced dermatologists managing fatigue-related diagnostic errors.¹ Although AI demonstrated strong performance, particularly for binary benign-versus-malignant classification, the findings suggest experienced dermatologists continue to provide the highest overall diagnostic accuracy when evaluating the complex and atypical lesions encountered in everyday clinical practice.

No regulatory action accompanies this diagnostic study, and the findings remain investigational pending validation in larger, more diverse patient populations.²

REFERENCES:

1. Anriot, J., Yan, S., Coste, C., Tschandl, P., Verlingue, L., Andremasse, C., Amini-Adle, M., Perrot, J. L., Ge, Z., Kittler, H., & Thomas, L. (2026). Limits of Artificial Intelligence Models for Skin Cancer Diagnosis in Realistic Settings. JAMA dermatology, 162(7), 701–708. https://doi.org/10.1001/jamadermatol.2026.149
2. Analysis: AI model outperforms early-career physicians in skin lesion diagnosis. Practical Dermatology. June 4, 2026. https://practicaldermatology.com/news/analysis-ai-model-outperforms-early-career-physicians-in-skin-lesion-diagnosis/2487332/
3. Yan, S., Yu, Z., Primiero, C. et al. A multimodal vision foundation model for clinical dermatology. Nat Med 31, 2691–2702 (2025). https://doi.org/10.1038/s41591-025-03747-y https://www.nature.com/articles/s41591-025-03747-y#citeas