Mitigating Bias in Skin Lesion Classification with Skin Tone Invariant Triplet Loss and Focal-Enhanced Disentangled Learning

Deep learning has been widely adopted in derma-tological classification. However, these models tend to amplify inherent biases present in training datasets, leading to disparities in classification performance between dark-skinned and light-skinned populations. This algorithmic bias significantly impedes the clinical deployment of dermatological diagnostic systems. Currently, the substantial computational complexity and memory footprint of conventional deep models restrict their practical implementation on resource-constrained mobile devices. In light of these challenges, we propose an improved and fair-enhanced lightweight skin disease classification network named CAS-ViT-SIRFDL. Through skin tone invariant triplet loss regularization and focal-enhanced disentangled learning, our approach improves group fairness without significantly increasing the number of parameters or compromising accuracy. Experimental results demonstrate that in the ISIC-PAD data set, which integrates data from ISIC2019 and PAD-UFES-20, our proposed method effectively mitigates model bias at a lower computational cost.