Explainable Deep Reinforcement Learning Framework for Automated Tuberculosis Detection and Clinical Decision Support

Abstract

Tuberculosis remains a major global health challenge, requiring accurate, rapid, and interpretable diagnostic support, particularly in resource-constrained settings. This study proposes an explainable deep reinforcement learning framework for automated tuberculosis detection and clinical decision support using chest X-ray images. The framework integrates image preprocessing, deep feature extraction, attention-based feature refinement, reinforcement learning-based decision-making, confidence estimation, explainability, and structured clinical recommendation generation. Radiological features extracted from preprocessed chest X-rays are transformed into compact state representations that guide a deep reinforcement learning agent toward appropriate diagnostic actions. A reward mechanism encourages correct tuberculosis classification, severity assessment, and clinically responsible handling of uncertain cases. Explainability is incorporated through attention visualization and Grad-CAM-based localization to highlight pulmonary regions contributing to the diagnostic outcome. The framework is evaluated using classification metrics, reinforcement learning behavior, ablation analysis, robustness testing, and generalization assessment. The analysis indicates that combining attention, reinforcement learning, confidence-aware prediction, and explainability can improve diagnostic stability and transparency compared with simpler configurations. The proposed system is intended to support, rather than replace, healthcare professionals by providing interpretable tuberculosis predictions, severity information, visual evidence, and decision-support recommendations. The framework offers a promising foundation for reliable artificial intelligence-assisted tuberculosis screening and clinical workflow integration. At scale.