A Digital Microscopy System and New Sputum Dataset for Deep Learning-based Tuberculosis Detection

Tuberculosis remains a major global health problem, particularly in low-resource settings where sputum smear microscopy is still widely used. This study introduces ID-SPUT30K, a large-scale microscopy dataset comprising 313 Ziehl-Neelsen-stained sputum slides and 31,300 high-resolution image patches acquired using a low-cost digital microscopy system. To enable reliable automated screening, we propose Mod-YOLOv10, a modified YOLOv10 architecture that integrates a RepVGG-based feature modeling module in the detection head for robust Acid-Fast Bacilli detection and IUATLD-based sputum grading. Extensive evaluations on a balanced blind test set demonstrate that Mod-YOLOv10 achieves 97.5% accuracy, 97.9% specificity, and consistently high sensitivity across all TB-positive grades. Cross-dataset validation on the MIDTI benchmark further confirms improved generalization under challenging staining and illumination conditions. These results indicate that the proposed dataset and model provide an effective and clinically relevant computer-aided screening framework to support routine tuberculosis diagnosis.