Study: Daneshjou 2022, Science Advances

Disparities in dermatology AI performance on a diverse, curated clinical image set

Citation

Daneshjou R, Vodrahalli K, Novoa RA et al (2022). Science Advances. PubMed id 35960806.

Open on PubMed

Cards that cite this study (1)

Each card gives its own wording for who was tested, how and what was found.

  • Finding a dermatologist who knows dark skin Tier B thin

    B 656 skin images (207 Fitzpatrick V-VI) in the Diverse Dermatology Images dataset; dermatologists and AI (n=656)

    Test of AI models and dermatologists on skin images by skin tone.

    AI models did worse on Fitzpatrick V-VI than I-II (for example ROC-AUC 0.55 against 0.64 for ModelDerm, with overlapping 95% intervals of 0.46 to 0.64 and 0.55 to 0.70; DeepDerm 0.50 against 0.61, HAM10000 0.57 against 0.72; figures from the full text, PMC9374341, opened 12 Nov 2026); dermatologists were also worse on dark skin; fine-tuning on the dataset closed the AI gap.

    As recorded in this card: Daneshjou R, Vodrahalli K, Novoa RA et al (2022), Science Advances, [primary]

    Note: From the research file (full text read 12 Nov 2026). Photos, not live clinic visits.

This is information, not medical advice. If you are worried, see a doctor or pharmacist. How we grade the evidence. All studies.