Explainable Tuberculosis Diagnosis from Chest X-rays Using Deep CNNs and LLaMA 3 Driven Conversational AI

Tuberculosis (TB) remains a major public-health challenge, particularly in settings where access to specialist radiological services and rapid diagnostic facilities is limited. Chest X-ray (CXR) imaging is widely used for pulmonary TB screening because it is comparatively accessible and can reveal radiographic abnormalities associated with the disease. However, many computer-aided detection systems primarily provide a classification score and offer limited information about the image regions influencing a prediction or the practical steps that should follow a screening result. This paper presents TB Care AI, an integrated artificial intelligence framework for explainable tuberculosis screening from chest X-ray images. The system combines an input-quality validation stage, a multi-task convolutional neural network (CNN), spatial segmentation, Gradient-weighted Class Activation Mapping (Grad-CAM), and a LLaMA 3 based conversational AI layer. The multi-task network performs diagnostic classification, pulmonary-zone or pathology prediction, and spatial segmenta-tion. A dual visual explanation mechanism combines segmenta-tion overlays and Grad-CAM attention maps. The conversational layer receives structured model outputs and transforms them into natural-language explanations and clinical decision-support information. The system also provides clinical report generation, temporary image handling, and healthcare-facility routing. On the reported evaluation benchmark, the framework achieved 98.40% accuracy, 97.60% precision, 98.80% sensitivity, 98.00% specificity, 98.20% F1-score, 98.56% F2-score, and a ROC-AUC of 0.992. The segmentation component achieved a Dice Similarity Coefficient of 91.40% and an Intersection over Union of 84.16%. The proposed platform is intended for screening and triage support and is not a substitute for formal clinical diagnosis or microbiological confirmation.