Towards an Unbiased Classification of Chest X-Ray Images Using a RL Powered ACGAN Framework

Abstract

Computer-aided diagnosis systems are invaluable tools for healthcare providers given the overwhelming volume of medical data at their disposal. However, a significant challenge of these systems is the existence of bias in their diagnostic outcomes, particularly affecting certain protected groups who are more susceptible to receiving incorrect diagnoses. In this thesis, we investigate bias mitigation strategies, leveraging the discriminator of auxiliary conditional generative adversarial networks (ACGANs) as well as reinforcement learning (RL) agents for the classification of chest X-ray images. Our research targets bias reduction by utilizing reward functions designed to diminish the true positivity rate disparity (TPR Gap) between males and females for the discriminator. We explore the impact of a hierarchical label distribution-based reward, a novel approach that aims to further improve bias in the diagnostic process. Through extensive evaluation and comparison, we study the disparities in precision, recall, f1 score, and true positivity rates across various different approaches. Also, we highlight the efficiency of each strategy in achieving more equitable diagnostic outcomes. The results show improvement in the TPR Gap when using the proposed method compared to the original classifier.