Addressing diagnostic resource imbalance in pulmonary tuberculosis detection from chest radiographs through cost-aware learning.

Automated early detection of pulmonary tuberculosis using chest radiographs is hampered by extreme class imbalance, resulting in poor performance across real-world settings where the clinical and operational cost of missed TB cases outstrips that of false positives. In this study, we address these challenges through the development of a cost-sensitive logistic regression framework that is explicitly tailored to better represent the highly uneven penalties associated with missing TB detection cases in screening scale clinical settings. By using a large chest X-ray dataset, our approach cascades cost-awareness from the training phase to the evaluation phase and learns clinically relevant trade-offs between costs due to false negatives (FNs) and FPs - simultaneously optimizing for multiple metrics including sensitivity, specificity. The results of our experiments show that the proposed method yields a significant improvement in detection sensitivity across risk and cost regimes (by approximately 60 %) at the same level of interpretability and operational scalability required for use in resource-limited or high-burden settings. We believe that these insights can help stakeholders develop more effective and equitable TB case finding in practice, particularly with adaptive risk-cost calibration for AI-driven TB screening tools.