Enhancing tuberculosis diagnosis: A deep learning-based framework for accurate detection and quantification of TB bacilli in microscopic images.

Abstract

Introduction: Tuberculosis (TB), a highly contagious disease, remains one of the leading causes of death globally. The proposed computer-assisted TB detection system enhances diagnostic accuracy and efficiency by integrating deep learning and segmentation techniques.

Materials and methods: It consists of two key subsystems: Automated field-ofview (FOV) recognition and TB bacilli segmentation. Using a motorized microscopic stage, the system systematically captures Ziehl-Neelsen-stained sputum smear images at 100x magnification. A customized Inception V3 model with transfer learning identifies FOVs containing TB bacilli, reducing variability and manual effort. Segmentation techniques, including coarse-level thresholding and shape descriptors like area, perimeter, and eccentricity, refine bacilli detection and eliminate artifacts.

Result: This study highlights the significant potential of deep learning and image processing techniques in advancing medical diagnostics, particularly TB detection. This framework has the potential to improve clinical outcomes and support global TB eradication efforts by providing a reliable tool for early TB diagnosis.

Conclusions: The system achieved a mean receiver operating characteristic score of 0.9505, a precision of 0.924, a recall of 0.882, and an F1 score of 0.902, demonstrating its potential to improve TB screening, particularly in resource-limited settings. By minimizing reliance on skilled technicians and enhancing diagnostic reliability, this approach offers a scalable solution for effective TB detection and severity assessment.