In the medical domain, accurate interpretation of chest X-ray (CXR) images is critical for diagnosis and treatment decisions. However, deep learning models trained on large datasets can be susceptible to spu- rious correlations, leading to erroneous interpretations and potentially harmful decisions. This study aims to address this issue in the CXR domain by proposing the use of reinforcement learning techniques and semi-supervised training. These methods actively select relevant CXR data samples while mitigating the influence of spurious correlations. The results demonstrate the effectiveness of these approaches in improv- ing prediction accuracy and decision-making performance compared to traditional data selection methods . This research contributes to the advancement of both technical state-of-the-art and clinical applications of deep learning in healthcare.