A Global Atlas of Digital Dermatology to Map Innovation and Disparities

TL;DR

SkinMap is presented, a multi-modal framework for the first comprehensive audit of the field's entire data basis and provides infrastructure to measure blind spots and steer strategic data acquisition toward undercovered regions of clinical space.

Abstract

The adoption of artificial intelligence in dermatology promises democratized access to healthcare, but model reliability depends on the quality and comprehensiveness of the data fueling these models. Despite rapid growth in publicly available dermatology images, the field lacks quantitative key performance indicators to measure whether new datasets expand clinical coverage or merely replicate what is already known. Here we present SkinMap, a multi-modal framework for the first comprehensive audit of the field's entire data basis. We unify the publicly available dermatology datasets into a single, queryable semantic atlas comprising more than 1.1 million images of skin conditions and quantify (i) informational novelty over time, (ii) dataset redundancy, and (iii) representation gaps across demographics and diagnoses. Despite exponential growth in dataset sizes, informational novelty across time has somewhat plateaued: Some clusters, such as common neoplasms on fair skin, are densely populated, while underrepresented skin types and many rare diseases remain unaddressed. We further identify structural gaps in coverage: Darker skin tones (Fitzpatrick V-VI) constitute only 5.8% of images and pediatric patients only 3.0%, while many rare diseases and phenotype combinations remain sparsely represented. SkinMap provides infrastructure to measure blind spots and steer strategic data acquisition toward undercovered regions of clinical space.

Figures (13)

• Figure 1: The SkinMap multi-modal framework for auditing the global dermatology landscape.a, Aggregation of 30 public datasets yielding 1.1 million unique images with sparse original metadata. b, Multi-modal training pipeline utilizing on imaging data and contrastive learning on image-text pairs generated via templated captions. c, Construction of a unified latent space by projecting diverse encoders into a shared low-dimensional manifold. d, Imputation engine training, where linear probes learn to predict missing attributes (e.g., fst, age, and sex) using partial meta-information. e, Applications of the resulting digital SkinMap atlas: (1) A clinician-facing search tool for finding semantically similar cases; (2) a structural audit tool for identifying dataset overlaps and homologies; and (3) a strategic acquisition guide for detecting underrepresented latent regions to direct future data collection.
• Figure 1: Embedding space density as an automated quality assurance mechanism.a, UMAP projection of the SkinMap embedding space overlaid with a Gaussian mixture model density map. Dashed circles indicate regions of hypersaturation (high density, $H1\!-\!H5$) and sparsity (low density, $L1\!-\!L4$). b, Boxplot distribution of log-density scores distinguishing core clusters from peripheral outliers. c, Representative samples from high-density peaks, revealing visual homogeneity and redundancy in common lesions. d, Representative samples from low-density regions, effectively isolating data quality failures, including non-clinical images, reconstruction artifacts, and heavy occlusions.
• Figure 2: Benchmarking imputation precision and metadata expansion.a, b, Radar plots comparing the predictive performance of the SkinMap ensemble against state-of-the-art foundation models MONET and PanDerm across multiple attributes on the internal validation set (a) and external hold-out datasets (b) and against random performance. The Ensemble model demonstrates superior performance in recovering missing demographic labels. c, Quantification of metadata coverage expansion achieved via imputation, showing significant gains for key attributes such as (+97.1 pp.) and geographic origin (+54.7 pp.). d, Comparative analysis of the model's imputation accuracy versus experienced dermatologists on a subset of 150 diverse cases.
• Figure 2: Validation of imputed demographic distributions. Comparison of demographic distributions derived from original sparse metadata (left column, a), the fully imputed SkinMap attributes (center column, b), and global population baselines (right column, c). The imputed distributions (center) effectively reconstruct the latent demographics of the dataset, revealing the true extent of exclusions (such as the scarcity of V–VI and pediatric data) that are otherwise obscured by missing metadata in the original archives.
• Figure 3: Systemic demographic and geographic disparities in the global dermatological datasets. Distributions of demographic attributes across the aggregated digital datasets (left, a) compared with global population statistics for dermatological visits (right, b). The comparison reveals sharp divergences: V--VI are marginalized (5.8% in datasets vs. 43.8% globally), pediatric populations are underrepresented (3.0% vs. 31.2%), and geographic origins are heavily skewed toward the Global North.