Lightweight Model Adaptation for Mitigating Bias in Deep Learning Models for Chest X-Ray Analysis

Clemence Mottez, Louisa Fay, Jean-Benoit Delbrouck, Curtis Langlotz

Keywords: Bias Mitigation, Chest X-ray, Convolutional Neural Network, eXtreme Gradient Boosting

TL;DR: This paper presents a lightweight hybrid approach combining CNN embeddings with an XGBoost classifier to mitigate biases related to sex, age, and race in chest X-ray diagnosis while maintaining high diagnostic performance.

Abstract: Deep learning (DL) models have demonstrated significant potential in improving chest X-ray (CXR) diagnosis. However, these models may exacerbate healthcare disparities. Addressing the inherent biases of DL models is essential to ensure their safe and reliable deployment in clinical practice. We suggest a novel bias mitigation approach that combines embeddings extracted by a Convolutional Neural Network (CNN) with an eXtreme Gradient Boosting (XGBoost) classifier. Our results show that this hybrid model significantly reduces bias across the sensitive attributes sex, age, and race, while maintaining comparable overall diagnostic performance and without the need for expensive model retraining. Our approach demonstrates that integrating simple, interpretable, and computationally efficient modifications into existing models can effectively enhance fairness in medical imaging.

Submission Number: 63