Disparities in Dermatology AI: Assessments Using Diverse Clinical Images

1. Introduction

• Public dermatology datasets lack key evaluation data: ISIC lacks clinical images, uncommon diseases, and diverse skin tones, while Fitzpatrick 17k can have unconfirmed malignancy labels.Most training and test images also come from private clinical datasets with sparse descriptors.Source: §1 · pp. 1–2
• DDI is a publicly accessible dataset of pathologically confirmed diagnoses, diverse skin tones, and uncommon diseases, enabling evaluation stratified by light and dark skin tones.The paper defines uncommon diseases here as those with incidence below 1 in 10,000.Source: §1 · p. 2
• The paper reports that existing algorithms lose performance on DDI, skin tone and disease rarity contribute to the drop-off, and robust training cannot correct biases without diverse training data.AI tools have been suggested for early skin cancer detection amid inadequate access to skin-disease care and a shortage of dermatologists in the United States.Source: §1 · pp. 1–2

2.1. Dataset design

• DDI images were retrospectively selected by reviewing Stanford Clinics pathology reports from 2010–2020.Diagnosis labeling, image quality control, and Fitzpatrick skin-type labeling are described in Appendix A.Source: §2.1 · p. 2

2.2. AI Algorithms

• Three algorithms—ModelDerm, DeepDerm, and HAM 10000—were tested for benign-versus-malignant image classification.ModelDerm outputs came from an online API in December 2020; HAM 10000 used the DeepDerm training methods but a different dataset.Source: §2.2 · p. 2
• DeepDerm was trained on mixed datasets with an 80/20 train-test split, image augmentation, balanced class sampling, binary cross-entropy loss, and Adam optimization.Its data sources included ISIC, two Stanford hospital datasets, the Edinburgh Dermofit Library, and open-access web data.Source: §2.2 · p. 2
• GroupDRO, CORAL, and CDANN variants used source-by-class groups as a proxy for skin-tone groups, with balanced sampling across 10 groups.The paper notes that FST metadata was unavailable for defining the groups directly.Source: §2.2 · p. 2

3. Results

• DDI comprised 440 images from 390 patients and was enriched for the lightest and darkest skin tones; photo-quality scores did not differ significantly between the compared tone groups.The reported photo-quality comparison was FST I-II versus FST IV-V, with p = 0.33.Source: §3 · pp. 2–3
• On DDI, ROC-AUC was 0.60 for ModelDerm, 0.53 for DeepDerm, and 0.65 for HAM 10000, below their reported original-test results.The original-test ROC-AUCs reported were 0.93–0.94 for ModelDerm, 0.88 for DeepDerm, and 0.92 for HAM 10000.Source: Table 1 · p. 3, §3 · p. 3
• Robust training methods modestly improved AUC, but did not significantly improve the performance gap between FST I-II and FST V-VI.The evaluation averaged results across five random seeds; Table 2 reports robust-training performance.Source: Table 2 · p. 3, §3 · p. 3
• Across DDI, ROC-AUC was higher for FST I-II than FST V-VI for all three algorithms; sensitivity also differed by skin tone for ModelDerm and DeepDerm.ModelDerm sensitivity was 0.41 versus 0.12 and DeepDerm sensitivity was 0.69 versus 0.23 for FST I-II versus FST V-VI, respectively.Source: §3 · pp. 3–4
• On DDI, ModelDerm and DeepDerm had lower malignancy sensitivity for FST V-VI than FST I-II, while HAM 10000 sensitivity was poor across subsets.HAM 10000 sensitivity was 0.04 on DDI and 0.03 on DDI-C, compared with 0.68 on its own test set.Source: §3 · p. 4
• Removing uncommon diseases improved overall ROC-AUC, but results remained below original-test performance and DeepDerm and HAM 10000 retained skin-tone performance gaps.In common-disease DDI, ROC-AUC was 0.70 for ModelDerm, 0.62 for DeepDerm, and 0.71 for HAM 10000.Source: §3 · p. 4

• Previously developed algorithms show performance drop-offs on DDI compared with ROC-AUC reported for their original test sets.DDI includes diverse skin tones and common and uncommon diseases, with every lesion confirmed pathologically.Source: §4 · p. 4
• Uncommon diseases and dark skin tones contribute to the drop-off; DeepDerm's shared clinical sites make site-specific differences less likely to explain disparities.The authors argue that dermatology AI training and testing datasets must include diverse data to avoid exacerbating existing disparities.Source: §4 · p. 4
• Training-data distributions influence performance; robust training methods did not improve disparity, suggesting the limitation lies in insufficiently diverse training data.ModelDerm, trained on a mix of Caucasian and Asian patients, did not show the DDI common-disease ROC-AUC drop-off between Fitzpatrick I-II and V-VI seen in DeepDerm and HAM 10000.Source: §4 · p. 4
• The study is limited by sample size, enrichment for FST I-II and V-VI, FST's incomplete coverage of skin tones, and the small number of testable algorithms.Expanding DDI is ongoing work.Source: §4 · p. 4

• Clinical images of biopsied lesions were retrieved from the electronic medical record and characterized by histopathologic diagnosis, benign or malignant status, biopsy date, and patient demographics.The Stanford Research Repository search identified self-described Black patients biopsied from January 1, 2010, onward; clinic-issued smartphones typically captured images from about six inches away.Source: Appendix A · p. 6
• Each FST V/VI lesion was matched to an FST I/II control by diagnosis, with demographic and biopsy-date criteria used where possible.When a strict diagnosis match was unavailable, a similar histopathologic diagnosis could be used to preserve the malignant-to-benign ratio.Source: Appendix A · p. 6
• Three independent board-certified dermatologists assigned image-quality scores using a previously developed scale.Source: Appendix A · p. 6
• FST labels were checked against clinical notes and, when available, demographic photos and lesion images; two dermatologists reviewed images and resolved discrepancies by consensus.FST I/II were categorized as light skin and FST V/VI as dark skin; FST III/IV were included in overall analyses but not enriched.Source: Appendix A · p. 6
• A board-certified dermatologist and dermatopathologist determined diagnosis and benign-versus-malignant labels from histopathologic diagnoses.Additional pathology tests were considered in borderline cases.Source: Appendix A · pp. 6–7

• Data will be shared through the ISIC repository, and algorithm code is provided via a Google Drive folder.Source: Appendix B · p. 7