Roxana Daneshjou

Who they are

On the 2022 paper her affiliations are the Department of Dermatology and the Department of Biomedical Data Science at Stanford School of Medicine. A Stanford Medicine article describes her as a practicing dermatologist and the lead author of the study.

This page records only what the paper and its published records state. Biography beyond the cited work is not given here.

Key papers

Disparities in dermatology AI performance on a diverse, curated clinical image set (2022), with Vodrahalli, Novoa and colleagues. The team built the Diverse Dermatology Images dataset, 656 pathologically confirmed images matched so that patients with light skin (Fitzpatrick types I and II) and dark skin (types V and VI) could be compared directly. They ran three published AI models on it. All three performed below what had been reported on their own test data. One model, DeepDerm, detected 69 per cent of malignancies on light skin and 23 per cent on dark skin. Dermatologists, who often label the images such models are trained on, also did worse on dark skin and on uncommon diseases. When two of the models were retrained on the new images, the gap between light and dark skin closed.

Key ideas

The study shows how a model can perform well on the patients it was developed on and fall short for others, a central example of algorithmic bias in medicine. Because retraining on more varied images closed the gap, the result points to the training data rather than to anything fixed in the method. It also shows the value of testing a tool on a carefully built outside dataset before trusting the accuracy figures its developers report.

---