Automatic Tuberculosis Classification in Sputum using YOLOv12 and IUATLD

Abstract

Tuberculosis remains a major global health challenge, particularly in resource-limited regions where rapid and accurate diagnosis is still difficult to achieve. This paper proposes an automated deep-learning framework for the detection, quantification, and classification of Acid-Fast Bacilli (AFB) in sputum smear images to address the limitations of conventional manual microscopy, which is labor-intensive, time-consuming, and prone to inter-observer variability. Our proposed method utilizes the YOLOv12 object detection architecture to identify tuberculosis bacilli from Ziehl–Neelsen-stained sputum images. We use a carefully annotated dataset prepared by expert microbiologists based on the International Union Against Tuberculosis and Lung Disease (IUATLD) grading standard. The experimental results show that the proposed model achieves strong detection performance with a mean Average Precision (mAP) of 84.7%, precision of 74.18%, recall of 73.90%, and F1-score of 74.04%. Furthermore, this paper maps the detected bacilli counts into standardized IUATLD diagnostic categories, including Scanty, 1+, 2+, and 3+, to ensure compatibility with routine clinical reporting procedures. The obtained results demonstrate that our proposed system can provide accurate, consistent, and efficient tuberculosis screening support. This study confirms that integrating YOLOv12 with IUATLD grading can produce a scalable and reliable automated diagnostic framework for improving TB microscopy analysis in high-burden laboratory environments.