Biasing & Debiasing based Approach Towards Fair Knowledge Transfer for Equitable Skin Analysis

TL;DR

The paper tackles fairness in CNN-based dermatology diagnosis by addressing biases across skin tones and gender. It introduces a two-biased-teachers knowledge-distillation framework in which each teacher is biased to a sensitive attribute and, through a multi-term KL-divergence loss, guides a student to learn fair representations without sacrificing accuracy. The approach uses a weighted loss L_total = λ L_CE + α L_bias_0 + β L_bias_1 + γ L_debias_0 + δ L_debias_1 and is evaluated on Fitzpatrick-17k and ISIC-2019, where it achieves superior fairness (lower Eopp0, Eopp1, Eodd) while maintaining or improving accuracy relative to state-of-the-art methods. The results are supported by extensive ablations and qualitative analyses, demonstrating robust fairness-accuracy trade-offs and suggesting practical deployment for equitable skin-analysis tools.

Abstract

Deep learning models, particularly Convolutional Neural Networks (CNNs), have demonstrated exceptional performance in diagnosing skin diseases, often outperforming dermatologists. However, they have also unveiled biases linked to specific demographic traits, notably concerning diverse skin tones or gender, prompting concerns regarding fairness and limiting their widespread deployment. Researchers are actively working to ensure fairness in AI-based solutions, but existing methods incur an accuracy loss when striving for fairness. To solve this issue, we propose a `two-biased teachers' (i.e., biased on different sensitive attributes) based approach to transfer fair knowledge into the student network. Our approach mitigates biases present in the student network without harming its predictive accuracy. In fact, in most cases, our approach improves the accuracy of the baseline model. To achieve this goal, we developed a weighted loss function comprising biasing and debiasing loss terms. We surpassed available state-of-the-art approaches to attain fairness and also improved the accuracy at the same time. The proposed approach has been evaluated and validated on two dermatology datasets using standard accuracy and fairness evaluation measures. We will make source code publicly available to foster reproducibility and future research.

Figures (7)

• Figure 1: The schematic architecture diagram of the proposed approach. We have shown how individual loss terms use logits from biased teachers to perform fair knowledge transfer (better view in color).
• Figure 2: Examples of clinical images representing skin disease conditions in the Fitzpatrick-17k dataset.
• Figure 3: Examples of dermoscopic images representing skin lesion classes in the ISIC-2019 dataset.
• Figure 4: Fairness comparison using bar plots (a), (b). We can see that the proposed approach ensures fairness across bias groups without sacrificing performance. Here, we used weight=1 for Bias Fair, Bias Dark, Debias Fair, Debias Dark.
• Figure 5: The bar plot obtained while employing different biasing and debiasing loss components at weight coefficient=0.6. We can observe that the proposed approach provides improvement over the baseline while ensuring fairness.