FairDD: Enhancing Fairness with domain-incremental learning in dermatological disease diagnosis

TL;DR

FairDD tackles fairness in dermatological disease diagnosis by adopting domain-incremental learning with a memory replay buffer to mitigate catastrophic forgetting. It integrates a cross-domain mixup module, supervised contrastive loss, a statistical parity disparity loss, and a distillation-based fine-tuning strategy, culminating in a unified objective that balances accuracy and demographic fairness. Experimental results on Fitzpatrick-17k and ISIC 2019 demonstrate strong fairness improvements and favorable trade-offs, indicating practical potential for unbiased dermatology AI. The work provides a concrete framework for achieving fairness-accuracy balance in domain-shift scenarios and highlights the importance of memory, data augmentation, contrastive learning, and distillation in fairness optimization.

Abstract

With the rapid advancement of deep learning technologies, artificial intelligence has become increasingly prevalent in the research and application of dermatological disease diagnosis. However, this data-driven approach often faces issues related to decision bias. Existing fairness enhancement techniques typically come at a substantial cost to accuracy. This study aims to achieve a better trade-off between accuracy and fairness in dermatological diagnostic models. To this end, we propose a novel fair dermatological diagnosis network, named FairDD, which leverages domain incremental learning to balance the learning of different groups by being sensitive to changes in data distribution. Additionally, we incorporate the mixup data augmentation technique and supervised contrastive learning to enhance the network's robustness and generalization. Experimental validation on two dermatological datasets demonstrates that our proposed method excels in both fairness criteria and the trade-off between fairness and performance.

Figures (4)

• Figure 1: The main framework of Network. At each stage, the student model is jointly trained on the input new domain samples and old domain samples from memory. During training, the teacher model provides knowledge to the student model while keeping its parameters fixed. After each stage of training, the student model’s parameters are copied to serve as the teacher model for the next stage. At the end of each epoch within a stage, the model undergoes distillation fine-tuning, and the memory is updated. Note that memory updates cease in the final phase. In the first stage, the student model is trained independently.
• Figure 2: Impact of distillation fine-tuning module.
• Figure 3: Impact of statistical disparity loss.
• Figure 4: Impact of buffer size.