Fairness and Explainability in Chest X-ray Image Classifiers

Abstract

Artificial intelligence (AI) is increasingly being used in healthcare, particularly for interpreting medical images. However, there are growing concerns regarding the presence of biases in these AI models, which raise important fairness considerations. This study investigates biases in artificial intelligence (AI) models for chest X-ray diagnosis and explores the role of Explainable AI (XAI) in understanding model decisions. Biases were observed in model performance across different patient groups and diseases. Various XAI techniques were employed to generate explanations for model decisions, and comparisons were made with explanations provided by doctors. We identified an optimized version of occlusion as the most accurate XAI technique in this case, which also provided a consistent accuracy of the explanations across all patient groups. Indeed, the explanations remained equally accurate regardless of variations in model performance for different subgroups, suggesting the absence of model bias amplification. Evaluating the correctness of XAI explanations posed challenges due to the limited availability of ground truth. In order to increase the power of our analysis, we explored alternative evaluation methods, like deletion or insertion curves, but reported them as unsuitable for chest X-ray images. We have therefore established some recommendations for using XAI on chest X-ray images. Given the reported absence of biases in the explanations, our aim is also to instill confidence in clinical stakeholders regarding XAI techniques.