Pulmonary tuberculosis (PTB) remains a major global health concern, particularly in high-burden countries such as Malaysia. Chest X-ray (CXR) is the most accessible screening modality, but its interpretation is subjective and prone to inter-observer variability. This study aimed to develop and evaluate a deep learning–based algorithm using YOLOv8 for automated PTB detection on CXR and to assess the impact of data augmentation on diagnostic performance. A retrospective cross-sectional study of 1,000 anonymized CXR images from Hospital Serdang was conducted, with CXRs categorized into three classes: highly suspicious PTB, low suspicion PTB, and no active lung lesion. Data were divided into training (70%), testing (20%), and validation (10%) sets, and pre-processing included DICOM-to-JPEG conversion, anonymization, and augmentation techniques such as flipping, rotation, cropping, brightness, and grayscale adjustments. Model performance was evaluated using accuracy, precision, recall, F1-score, ROC and Precision–Recall (PR) curves, and Grad-CAM visualization for explainability. Without augmentation, YOLOv8 achieved 74% accuracy (precision 0.78, recall 0.74), while with augmentation, accuracy improved to 85%, recall for highly suspicious PTB increased from 0.75 to 0.87, and macro F1-score rose from 0.76 to 0.85. ROC analysis demonstrated macro-average AUC improvement from 0.71 to 0.85, and PR curves showed micro-average AP rising from 0.55 to 0.74. Grad-CAM highlighted radiologically relevant areas such as upper-lobe consolidation and cavitary lesions, supporting clinical interpretability. Overall, the YOLOv8 model achieved diagnostic performance comparable to commercial CAD systems such as CAD4TB and Lunit INSIGHT, with data augmentation significantly enhancing sensitivity and generalization. This locally trained AI model demonstrates strong potential for scalable PTB screening, particularly in resource-limited settings.

Keywords: Artificial Intelligence; Deep Learning; YOLOv8; Pulmonary Tuberculosis; Chest X-Ray