Source-linked AI summary
When Annotation Patterns Are Learned: Skin-Tone Dependence in Dermatology Concepts
Y. Li, M. Bigliardi-Qi, S. Bedros, P. Bigliardi, N. Papanikolopoulos
TL;DR
Concept bottleneck models rely on reliable clinical-concept annotations, yet how consistently dermatology concepts are annotated across skin tones is uncertain. The paper audits SkinCon and tests whether a trained concept model and label-error detectors reflect or identify tone-dependent patterns. Erythema and pigmentary annotations differ by tone within diseases, the model reproduces the pattern, and three detectors trail a disease-only baseline on potential erythema omissions.
Problem
Whether dermatology concept annotations are reliable across skin tones matters because concept bottleneck models rely on them for inspectable and correctable diagnoses.
Source: §1 · p. 1Method
The paper audits SkinCon annotations within diseases, tests an evidential concept model for inherited patterns, and compares three detectors with a disease-only baseline on potential erythema omissions.
Source: §2.1 · p. 1, §2.3 · p. 2Results
Erythema annotation odds are lower on dark skin (OR 0.17), while hyperpigmentation (OR 6.19) and hypopigmentation (OR 4.50) odds are higher; the model reproduces the pattern and three detectors trail the baseline.
Source: §3.1 · p. 2, §5 · p. 4Takeaways & Limitations
Concept annotations should be audited within each subgroup, with independent review of suspected errors.
Source: §5 · p. 4The SkinCAP check cannot resolve whether pigmentary differences reflect annotation or real prevalence because dermatologists saw the same photographs and may have missed the same signs.
Source: §4 · p. 4
Abstract
from arXiv · showhide
Concept bottleneck models (CBMs) let clinicians inspect and correct skin disease diagnoses through clinical concepts, but only if the concept annotations are reliable. We therefore audit the reliability of SkinCon, a widely used dermatology concept dataset, across skin tones. Within the same disease and with near duplicates removed, erythema is annotated far less often on dark than on light skin, with an odds ratio of 0.17. In contrast, hyperpigmentation and hypopigmentation are annotated more often, with odds ratios of 6.2 and 4.5, respectively. We refer to this pattern as tone-dependent concept annotation. An evidential concept model trained on these annotations learns the same pattern and reproduces it on unseen images. Such a model may agree with potential erythema omissions rather than contradict them. Consistent with this, three label error detectors underperform a disease-only baseline at finding erythema that dermatologist captions mention but SkinCon labels as absent. The three detectors are based on confident learning, training dynamics, and the model's own evidential uncertainty. Concept annotations should thus be audited within each subgroup before explanations built on them can be trusted.
1. INTRODUCTION
Skin-tone differences in visibility can shape dermatology annotations, while CBMs depend on those annotations to make their reasoning readable and correctable. The paper audits tone-dependent annotation and whether models and error detectors reproduce or detect it.
1. INTRODUCTION
- On darkly pigmented skin, reduced color contrast and surrounding pigment can make erythema harder to see, including in photographs.The British Association of Dermatologists issued guidance on describing erythema in skin of color.Source: §1 · p. 1
- CBMs route diagnoses through expert-defined clinical concepts that clinicians can inspect and correct, but both properties assume reliable concept annotations.Systematic subgroup misannotation can leave explanations faithful to the model but grounded in incorrect labels, while clinicians may also miss the error in the photograph.Source: §1 · p. 1
- Detectors trained on the labels they assess may reproduce visibility-related omissions rather than disagree with them, a case that violates confident learning’s noise assumption.The paper notes that group-dependent noise had been modeled for fair classification, but not for concept annotation.Source: §1 · p. 1
- SkinCon annotates erythema less often and pigmentary changes more often on dark skin within the same disease; a trained model reproduces the pattern, and three detectors trail a disease-only baseline.The paper calls this pattern tone-dependent concept annotation.Source: §1 · p. 1
2. METHODS
The methods audit concept annotation across skin tones, test whether a trained model inherits annotation patterns, and evaluate three label-error signals against a disease-only baseline.
2.1. Data
- The primary analysis uses 2,217 deduplicated SkinCon-annotated images, with light, mid, and dark groups; a full set and alternate Fitzpatrick17k tone labels provide checks.Fitzpatrick17k has 16,577 images, and SkinCon annotations describe images using 48 binary concepts; 22 concepts with more than 50 positive examples are retained.Source: §2.1 · pp. 1–2
2.2. Quantifying annotation dependence on skin tone
- Disease-fixed-effect logistic regression estimates each concept’s dark-versus-light annotation odds ratio, using diseases with both tone groups and annotation values and bootstrap intervals.The model shares the dark-skin log-odds shift across diseases, so differences in disease composition do not affect the within-disease ratio.Source: §2.2 · p. 2
2.3. Detection protocol and baseline
- Potential erythema omissions are images labeled absent by SkinCon whose checked SkinCAP captions mention erythema about the image itself; every signal is compared with a disease-only AUC baseline.The baseline uses SkinCon annotations but no image information or SkinCAP reference; caption sentences without a mention count as negatives, though the concept may still be present.Source: §2.3 · p. 2
2.4. Signals under test
- The tested signals are confident learning’s self-confidence, training-dynamics AUM, and evidential uncertainty; low confidence or AUM and high uncertainty indicate suspected errors.The signals use evi-CEM predictions from a model trained with five-fold cross-validation and three seeds, separately on deduplicated and full annotated sets.Source: §2.4 · p. 2
3. RESULTS
The three tested erythema detectors trail a disease-only baseline, while SkinCon annotations and a model trained on them show tone-dependent patterns. Confident-learning flags also follow annotation differences rather than clearly surfacing potential erythema omissions.
3.1. Annotation depends on skin tone in both directions
- 0.17 for erythema, 6.19 for hyperpigmentation, and 4.50 for hypopigmentation are the deduplicated-set odds ratios for annotations on dark versus light skin.The pattern persists in the full set and with Centaur Labs labels; the SkinCAP check did not confirm more caption mentions of erythema on dark skin (OR 0.93 [0.16, 1.81], 109 images across seven diseases).Source: §3.1 · pp. 2–3, Figure 1 · p. 3
3.3. Three detectors trail the baseline on erythema
- All three erythema detectors trail the disease-only baseline; u reaches AUC 0.48, and paired differences are −0.24 (u), −0.08 (AUM), and −0.09 (CL).The result holds with a five-fold cross-validation baseline, whose AUC is 0.718; all three full-set erythema estimates also fall below baseline, while tone-group estimates are small and sometimes have intervals crossing zero.Source: Table 1 · p. 3, §3.3 · pp. 3–4
3.2. The model inherits the pattern
- 0.25 [0.17, 0.32] is the model’s erythema odds ratio on unseen images, compared with 0.17 for annotations; thresholding yields ratios of 0.17, 8.94, and 4.37 for the three concepts.The model reproduces the annotation pattern for erythema and both pigmentary concepts; the thresholded effects are about as strong as, or stronger than, those in annotations.Source: §3.2 · p. 3
3.4. Detector flags follow the annotation pattern
- Confident-learning flags rank the two pigmentary concepts near the top and significant after correction, while erythema is never significant and would be overlooked.Erythema marked present is flagged 2.98–4.33 times as often on dark skin, but erythema marked absent only 0.41–0.55 times as often; the flags follow the annotation pattern in either case.Source: Table 2 · p. 3, §3.4 · p. 4
4. DISCUSSION
The discussion interprets erythema’s annotation pattern as plausibly related to visibility, while pigmentary patterns may partly reflect real differences. It argues that models and detectors can inherit annotation patterns and outlines subgroup audits alongside limitations of the evidence.
Interpretation
- Erythema may be less visible on dark skin, whereas pigmentary changes may stand out more and may partly reflect real differences in prevalence.The authors note that pigmentary changes can be more expected and more common after inflammation in darker skin; the SkinCAP check cannot resolve whether differences are real or annotation-related because dermatologists saw the same photographs.Source: §4 · p. 4
Why detectors may miss omissions
- Image-dependent omissions may be learned from faint images, allowing the model to agree with labels that disagreement-based detectors would otherwise flag.Evidential uncertainty measures the model’s evidence rather than whether it disagrees with the label, and faint erythema can also be missed on light skin.Source: §4 · p. 4
Implications
- The authors recommend disease-adjusted odds ratios by tone, comparison with a disease-only baseline, and annotation guidance describing erythema on dark skin.They note that erythema may appear violaceous or dusky rather than red on dark skin.Source: §4 · p. 4
Limitations
- The audit identifies concepts annotated differently, not which images are mislabeled; evidence comes from one dataset with few dark-skin images and one concept model.Other stated boundaries include wide intervals, skin tone measured with a noisy proxy, and a reference that may favor the disease-only baseline because diagnosis could influence captions.Source: §4 · p. 4
5. CONCLUSION
SkinCon shows tone-dependent concept annotation, and a trained concept model reproduces the pattern on unseen images. Three detectors trail a disease-only baseline in finding potential erythema omissions.
5. CONCLUSION
- Erythema labels are less frequent on dark skin within the same disease, with an odds ratio of 0.17.Pigmentary labels are more frequent on dark skin.Source: §5 · p. 4
- The concept model inherits the annotation pattern, with an erythema odds ratio of 0.25.The model reproduces the pattern on unseen images.Source: §5 · p. 4
- Three detectors trail a disease-only baseline in finding potential erythema omissions.The conclusion recommends subgroup audits and independent review of suspected errors.Source: §5 · p. 4
6. COMPLIANCE WITH ETHICAL STANDARDS
The study retrospectively analyzes open-access data from Fitzpatrick17k, SkinCon, and SkinCAP.