Artificial intelligence promises democratized access to dermatological care, but model reliability depends on the quality and comprehensiveness of the underlying data. Despite rapid growth in publicly available dermatology images, the field lacks quantitative indicators to measure whether new datasets expand clinical coverage or merely replicate what is already known. Existing collections are widely assumed to be biased, yet the scale and topography of these gaps have remained unquantified, leaving the community without a road map for strategic data acquisition. Here, we show, by unifying 1.1 million images from 29 public datasets into a queryable semantic atlas, that the global archive is markedly skewed and increasingly redundant: darker skin tones (Fitzpatrick V–VI) constitute only 11.0% of images, pediatric patients only 2.3%, and informational novelty has plateaued despite exponential growth, revealing where targeted data acquisition is most needed.
We thank Martin A. Walter for valuable suggestions that helped improve the clarity of the manuscript.
S.L., L.A. and M.P. are supported by the Swiss National Science Foundation (SNSF) under grant 20HW-1_228541. Open access funding provided by University of Basel.
Department of Biomedical Engineering, University of Basel, Hegenheimermattweg 167b, Allschwil, Basel, Switzerland
Source: Nature - Machine Learning — Read the full article on the original site ↗
