PaperlayerView exact revision on arXiv

Source-linked AI summary

Disparities in Dermatology AI Performance on a Diverse, Curated Clinical Image Set

Roxana Daneshjou, Kailas Vodrahalli, Roberto A Novoa, Melissa Jenkins, Weixin Liang, Veronica Rotemberg, Justin Ko, Susan M Swetter, Elizabeth E Bailey, Olivier Gevaert, Pritam Mukherjee, Michelle Phung, Kiana Yekrang, Bradley Fong, Rachna Sahasrabudhe, Johan A. C. Allerup, Utako Okata-Karigane, James Zou, Albert Chiou

arXiv:2203.08807v1eess.IVcs.AIcs.CVcs.LG

TL;DR

Dermatology AI models have been insufficiently assessed on diverse skin tones and uncommon diseases. Using the pathologically confirmed DDI dataset, this study found worse performance on dark skin and uncommon diseases, while fine-tuning on diverse images closed skin-tone performance gaps.

Abstract

from arXiv · show

Access to dermatological care is a major issue, with an estimated 3 billion people lacking access to care globally. Artificial intelligence (AI) may aid in triaging skin diseases. However, most AI models have not been rigorously assessed on images of diverse skin tones or uncommon diseases. To ascertain potential biases in algorithm performance in this context, we curated the Diverse Dermatology Images (DDI) dataset-the first publicly available, expertly curated, and pathologically confirmed image dataset with diverse skin tones. Using this dataset of 656 images, we show that state-of-the-art dermatology AI models perform substantially worse on DDI, with receiver operator curve area under the curve (ROC-AUC) dropping by 27-36 percent compared to the models' original test results. All the models performed worse on dark skin tones and uncommon diseases, which are represented in the DDI dataset. Additionally, we find that dermatologists, who typically provide visual labels for AI training and test datasets, also perform worse on images of dark skin tones and uncommon diseases compared to ground truth biopsy annotations. Finally, fine-tuning AI models on the well-characterized and diverse DDI images closed the performance gap between light and dark skin tones. Moreover, algorithms fine-tuned on diverse skin tones outperformed dermatologists on identifying malignancy on images of dark skin tones. Our findings identify important weaknesses and biases in dermatology AI that need to be addressed to ensure reliable application to diverse patients and diseases.

Results

On the biopsy-confirmed, diverse DDI benchmark, dermatology AI models performed worse overall, particularly on dark skin tones and uncommon diseases. Fine-tuning on DDI narrowed skin-tone performance gaps, while dermatologist visual labels also showed differential noise.

Dataset

  • The DDI dataset contains pathologically confirmed lesions with diverse skin tones, enabling matched comparisons between FST I-II and FST V-VI images.FST categories were assigned through chart review and consensus review by two board-certified dermatologists.Source: Results · p. 2

Overall performance

  • ROC-AUC fell on DDI to 0.65 for ModelDerm, 0.56 for DeepDerm, and 0.67 for HAM 10000, below each model’s original test performance.Original reported ROC-AUCs were 0.93-0.94, 0.88, and 0.92, respectively.Source: Previously developed Dermatology AI… · p. 3

Skin-tone disparities

  • ROC-AUC was lower on FST V-VI than FST I-II for ModelDerm, DeepDerm, and HAM 10000: 0.55 versus 0.64, 0.50 versus 0.61, and 0.57 versus 0.72.The subsets were matched for diagnostic class and patient demographics.Source: Previously developed Dermatology AI… · p. 3
  • Sensitivity for malignancy detection was lower on FST V-VI than FST I-II for ModelDerm, 0.12 versus 0.41, and DeepDerm, 0.23 versus 0.69.Both differences were statistically significant; HAM 10000 had poor sensitivity across all DDI subsets.Source: Previously developed Dermatology AI… · p. 3

Disease distribution

  • Removing uncommon diseases improved overall ROC-AUC, but performance remained below the models’ original test results and disparities persisted for DeepDerm and HAM 10000.On common diseases, ROC-AUC was 0.74 for ModelDerm, 0.64 for DeepDerm, and 0.71 for HAM 10000.Source: Previously developed Dermatology AI… · p. 3

Implications and scope

  • DDI provides a public benchmark intended to evaluate and reduce performance disparities, while its biopsy-curated composition enriches the dataset for ambiguous lesions requiring clinical biopsy.The dataset is tailored toward triaging malignant versus benign lesions rather than comprehensive dermatology diagnosis.Source: Discussion · pp. 4–6

Dermatologist labeling

  • Dermatologist consensus labels were less accurate against biopsy-proven labels for uncommon diseases and FST V-VI images, indicating differential label noise across skin tones.The authors attribute this as a potential issue in visual-consensus labels used for AI datasets, without making claims about general dermatologists’ overall performance.Source: Fine-tuning on diverse data can close… · p. 4, Discussion · p. 5

Materials and Methods

The study assembled and analyzed a biopsy-grounded, diverse dermatology image benchmark alongside existing AI algorithms, dermatologist labels, and fairness-aware training methods. Analyses covered image quality, disease frequency, skin-tone matching, model outputs, and fine-tuning on DDI.

Dataset and analysis

  • FST I/II lesions were matched to FST V/VI lesions by identical histopathologic diagnosis, diagnosis period, sex, age, and other demographic controls.Source: Materials and Methods · p. 6

Reader study

  • Three independent board-certified dermatologists assessed each image without access to image data or pathology labels, and 13 images failing quality metrics were removed.Source: Materials and Methods · p. 7

Fairness-aware training

Dataset and analysis

  • 656 clinical images were analyzed, with uncommon diagnoses identified using incidence below 1 in 10,000 or expert dermatologist and dermatopathologist judgment.Source: Materials and Methods · p. 8

Performance analysis

  • Model predictions used algorithm-generated probabilities for ROC-AUC, while sensitivity and specificity used development-set cutoffs and dermatologist majority voting defined ensemble labels.Source: Materials and Methods · p. 8

Model training

  • DeepDerm and HAM 10000 were fine-tuned on DDI using randomized 60-20-20 train-validation-test splits, augmentation, Mixup, Adam optimization, and repeated random seeds.Source: Materials and Methods · p. 9

Supplemental Tables and Figures

The supplemental materials organize DDI image counts and performance comparisons across algorithms, dermatologist ensembles, disease-frequency subsets, and skin-tone strata. They also report DeepDerm performance after fairness-aware training.

Supplemental Table 1

  • Supplemental Table 1 reports image counts after quality filtering, with parenthetical counts identifying common-diagnosis subsets.Source: Table 1 · p. 12

Supplemental Table 2

  • Supplemental Table 2 compares ModelDerm, DeepDerm, HAM 10000, and dermatologist ensembles across all 656 images, 564 common-diagnosis images, and light versus dark skin tones.Source: Table 2 · p. 12
  • Supplemental Table 2 reports ROC-AUC, sensitivity, and specificity, with algorithm ROC-AUC calculated from each model’s probability outputs.Source: Table 2 · p. 12, Supplemental Tables and Figures · p. 13

Supplemental Table 3

  • Supplemental Table 3 reports DeepDerm performance overall and separately for FST I-II and FST V-VI after fairness-aware training.Source: Table 3 · p. 13

Related papers

Loading 2203.08807v1…