A large dermatology image collection can contain many examples of a common condition and few examples from particular groups. Gröger and colleagues assembled 29 public collections into one atlas to ask where that unevenness occurs. A separate, earlier experiment tested three models on a deliberately curated set of skin images. The two papers let readers compare what is available to build models with what happens when a model is tested.
- The atlas pooled approximately 1.1 million images from 29 public datasets.
- Its abstract reports 11.0% for images of Fitzpatrick types V–VI and 2.3% paediatric representation.
- In a separate 2022 study, one model’s ROC-AUC was 0.72 on types I–II images and 0.57 on types V–VI images from the DDI dataset.
- Paper date
- 21 November 2026
- Input datasets
- 29 publicly available collections
- Pooled scale
- Approximately 1.1 million images
- Research question
- Coverage and redundancy in the collected images
Original publication: 21 Nov 2026 · The date above refers to this brief.
Two representation measures reported in the atlas abstract
percent
From counting images to testing a model
Count what is available
Record images, groups and the origins of the labels.
Build a test comparison
Define the population and how each diagnosis is confirmed.
Read performance and uncertainty
Report subgroup results with intervals and consider where the model was tested.
An image count describes coverage of a collection
The atlas pooled dermatology images for a study of what existing public collections cover. The abstract reports 11.0% for darker skin types V–VI on the Fitzpatrick scale and 2.3% for paediatric representation. These are separate summaries. To see coverage of children with darker skin, a reader would need the joint counts and information about how age and skin type were assigned when labels were missing. The abstract also reports that the authors’ measure of novelty grew more slowly than the total archive, raising a different question about overlap.
Source 1 ↗Coverage can be engineered into a test set
Daneshjou and colleagues built the earlier DDI test set from 656 images tied to pathology-confirmed diagnoses. It included 208 images of types I–II, 241 of types III–IV and 207 of types V–VI. The last category accounts for 31.6% of the test set. The type I–II and V–VI subsets were matched on selected characteristics, including whether lesions were benign or malignant. The curated proportions reflect what the investigators needed for their comparison.
Source 2 ↗Model evaluation has its own outcome
On the DDI test set, the classifier trained on HAM10000 had an ROC-AUC of 0.72 for types I–II and 0.57 for types V–VI. The point estimates differ by 0.15 on a 0-to-1 scale. ROC-AUC measures ranking across possible cutoffs. The estimates apply to this model and evaluation set, and each has uncertainty. The atlas percentages describe representation. A different model or test population could produce different performance results.
Source 1 ↗Source 2 ↗Three questions for the next image benchmark
First, find out how many people are represented and whether several images came from the same person. Next, look for subgroup and intersection counts, along with the source and completeness of the labels. Finally, look for evaluation on images collected independently of model development. Those checks would help show whether a larger pool brings useful coverage and whether an imaging model performs consistently for people represented in smaller groups.
Source 1 ↗Source 2 ↗Check representation before interpreting a performance claim
Publisher’s abstract and publication record for the accepted article. Checked the scale and headline representation figures. The detailed annotation and overlap analysis were outside this source check.
01What does the atlas count?
- What was observed
- The analysis pooled approximately 1.1 million images from 29 public datasets.
- Where the conclusion stops
- Image inventory reflects the included datasets. It is distinct from patient counts and model outcomes.
Source 1 · Published abstract
02What did the separate DDI experiment find?
- What was observed
- One classifier’s ROC-AUC differed between types I–II and V–VI subsets.
- Where the conclusion stops
- The point estimates depend on the test dataset and classifier.
Source 2 · Daneshjou et al., Figure 1B and 1C
Numbers you can inspect
| Measure | Value & unit | Origin & method |
|---|---|---|
| Public datasets pooled | 29 datasets | Reported Number of public image collections described. Source 1 · Published abstract |
| Approximate pooled size | 1100000 images | Reported Approximate count reported as 1.1 million. Source 1 · Published abstract |
| Fitzpatrick types V–VI share in the atlas | 11 percent | Reported Reported image representation. Source 1 · Published abstract |
| Paediatric representation in the atlas | 2.3 percent | Reported Reported representation figure. Source 1 · Published abstract |
| DDI images | 656 images | Reported Count after quality exclusions. Source 2 · Daneshjou et al., Results and Methods |
| DDI types V–VI images | 207 images | Reported Subgroup count in the curated test set. Source 2 · Daneshjou et al., Results and Table S1 |
| DDI types V–VI share | 31.6 percent | Calculated Divide 207 by 656 and multiply by 100. Source 2 · Daneshjou et al., Results |
| HAM10000 classifier on DDI types I–II images | 0.72 ROC-AUC | Reported 95% CI 0.63 to 0.79. Source 2 · Daneshjou et al., Figure 1B |
| HAM10000 classifier on DDI types V–VI images | 0.57 ROC-AUC | Reported 95% CI 0.48 to 0.67. Source 2 · Daneshjou et al., Figure 1C |
| Difference between the two point estimates | 0.15 ROC-AUC units | Calculated Subtract 0.57 from 0.72. Source 2 · Daneshjou et al., Figure 1 |
Compare the actual experiments
These studies answer different questions. Read the unit and endpoint before comparing results.
| Study | Unit & setting | Readout | Interpretation boundary |
|---|---|---|---|
| Pooled image atlas and curated performance evaluation Source 2 · Daneshjou et al., Results | 29 public datasets and one DDI test set | Representation and ROC-AUC | Read each result against its dataset and what was measured. |
| Benchmark transfer across groups Source 2 · Daneshjou et al., Figure 1 | Test sets with their own subgroup composition | Model ranking in specified test subsets | A headline aggregate may conceal group differences. |
The export includes claims, available numbers, methods and source locations. It contains our reading notes and published summaries; it is not raw participant data or an independent reanalysis.
Evidence update · 22 Nov 2026
First publication. Any calculations in this brief use reported figures. The reading notes identify possible follow-up questions. AI source check; no human editorial or clinical review.
Image representation and model performance
| Study and unit | Reported measure | How to read it |
|---|---|---|
| Atlas, pooled archive | About 1.1 million images from 29 datasets | Coverage spans the included public collections. |
| Atlas, subgroup representation | 11.0% types V–VI and 2.3% paediatric representation | Separate categories. Joint counts require a separate breakdown. |
| DDI, HAM10000 classifier | ROC-AUC 0.72 (95% CI 0.63 to 0.79) for types I–II | ROC-AUC 0.57 (95% CI 0.48 to 0.67) for types V–VI. |
The atlas abstract and DDI Figure 1 answer separate questions.
Your questions, answered
Does 1.1 million mean 1.1 million people?
The figure counts images. Several images can come from one person, so patient counts need a separate field.
Can the two atlas percentages tell us the share of children with darker skin?
A joint breakdown by age and skin type would answer that question. The two headline proportions describe separate summaries.
What does the 0.15 difference measure?
It is the difference between two reported ROC-AUC point estimates for the same classifier on two subsets of DDI.
How would readers check a new model rather than this earlier one?
Look for an independent evaluation with comparable labels, reported group counts and uncertainty intervals.
Limits of this interpretation
- The new atlas estimates are drawn from the publisher’s abstract. Detailed methods, label provenance and uncertainty for those estimates were outside this check.
- The collections represent public datasets selected for the atlas. Their proportions do not estimate how often conditions occur in clinical practice.
- The DDI experiment assessed specific older models on its own curated set. Its point estimates should be read with their confidence intervals.
Sources & transparency
- Gröger, Lionetti, Gottfrois et al. (2026) — A global atlas of digital dermatology to map innovation and disparities
Publisher’s abstract and publication record for the accepted article. Checked the scale and headline representation figures. The detailed annotation and overlap analysis were outside this source check. · Accessed 22 Nov 2026
DOI: 10.1038/s41467-026-78602-w - Daneshjou et al. (2022) — Disparities in dermatology AI performance on a diverse, curated clinical image set
Full paper. Checked the curated test-set counts, the grouping approach and ROC-AUC results for the HAM10000 classifier on DDI subsets. · Accessed 22 Nov 2026
DOI: 10.1126/sciadv.abq6147
Prepared and source-checked with AI for Press-news Team. The source notes identify the material available during preparation. The article received no human editorial or clinical review. The source list supports the reported results and the calculations shown.
Source check: AI source check — primary research and stated source limits
Clinical review: Not applicable to this educational guide
Suggest a correction