[Paper Review] LesionTABE: Equitable AI for Skin Lesion Detection

Rocio Mexia Diaz, Yasmin Greenway+1 more|arXiv (Cornell University)|Jan 6, 2026

Cutaneous Melanoma Detection and Management0 citations

TL;DR

LesionTABE couples adversarial debiasing with dermatology-focused foundation model embeddings to improve fairness across skin tones in skin lesion detection, achieving notable gains in both fairness and diagnostic performance.

ABSTRACT

Bias remains a major barrier to the clinical adoption of AI in dermatology, as diagnostic models underperform on darker skin tones. We present LesionTABE, a fairness-centric framework that couples adversarial debiasing with dermatology-specific foundation model embeddings. Evaluated across multiple datasets covering both malignant and inflammatory conditions, LesionTABE achieves over a 25\% improvement in fairness metrics compared to a ResNet-152 baseline, outperforming existing debiasing methods while simultaneously enhancing overall diagnostic accuracy. These results highlight the potential of foundation model debiasing as a step towards equitable clinical AI adoption.

Motivation & Objective

• Address bias in dermatology AI due to under-representation of darker skin tones in training data.
• Develop a fairness-centric framework that preserves lesion-relevant features while removing skin-tone information.
• Evaluate across malignant and inflammatory skin conditions with external validation datasets.
• Assess the benefit of dermatology-specific foundation model embeddings for fairness and accuracy.

Proposed method

• Combine adversarial debiasing with tone-invariant learning to suppress skin-tone information while retaining lesion features.
• Evaluate three debiasing architectures (TABE, VAE with adaptive resampling, FairDisCo) with a ResNet-152 backbone as baseline.
• Incorporate dermatology-specific foundation model embeddings using LesionCLIP as an alternative feature extractor.
• Train on Fitzpatrick17k and test externally on PAD-UFES (cancer) and SCIN (eczema/psoriasis) to assess external validity.
• Use balanced accuracy as primary performance metric and Equality of Opportunity (EOM) and Predictive Quality Disparity (PQD) as fairness metrics.
• Provide publicly available code and data handling scripts for replication.

Experimental results

Research questions

• RQ1 Can adversarial debiasing combined with foundation-model embeddings reduce skin-tone related disparities in malignant lesion detection without sacrificing accuracy?
• RQ2 How do different debiasing architectures compare internally and externally when evaluated on skin-tone diverse datasets?
• RQ3 Does a dermatology-focused foundation model (LesionCLIP) improve fairness and generalization across tasks (malignant vs inflammatory lesions) compared to ImageNet-based features?

Key findings

• LesionTABE (TABE + LesionCLIP) achieved the best trade-off with a balanced accuracy of 71.8% and an EOM of 0.56, representing relative improvements of 34% in fairness and 5% in performance over the baseline.
• On malignant lesion detection, TABE with LesionCLIP consistently yielded higher fairness in internal and external evaluations compared to other combinations.
• For eczema vs psoriasis classification, LesionTABE achieved a balanced accuracy of 62.8% and a PQD of 0.80 with relative gains of 25% in fairness and 8% in balanced accuracy over the baseline.
• External validation highlighted that some methods (e.g., FairDisCo) excelled internally but underperformed externally, underscoring the importance of external validation for debiasing strategies.
• Foundation-model embeddings from LesionCLIP provided robust fairness and generalization across skin tones despite training on predominantly lighter skin types.