Assessing Bias in Skin Lesion Classifiers with Contemporary Deep Learning and Post-Hoc Explainability Techniques

Abstract

As Artificial Intelligence (AI) is increasingly utilized in dermatology, ensuring fairness in the development of Machine Learning models is crucial, particularly in skin lesion classification, where decisions can significantly impact people's lives. This study investigates the presence of biases between different Fitzpatrick Skin Types in baseline pretrained models and evaluates various training techniques to mitigate these disparities. An unsupervised skin transformer is developed to adjust an image's Fitzpatrick Skin Type (FST), and joint regularization and synthetic image blending methods are employed to address bias concerns. Additionally, eXplainable Artificial Intelligence (XAI) techniques, such as Gradient-weighted Class Activation Mapping (Grad-CAM), are utilized to identify any underlying reasons for bias in the models. The results indicate that joint regularization and synthetic blending methods enhance the area under the curve performance and fairness. Meanwhile, XAI was found to be a valuable tool for fine-tuning Deep Learning models and uncovering problems. These findings can aid in developing accurate and unbiased skin lesion classification models, promoting equitable healthcare, and improving patient outcomes.