Automated tuberculosis diagnosis: A hybrid approach using attention-residual U-Net segmentation with ensemble classification

Tuberculosis (TB), caused by Mycobacterium tuberculosis, remains a critical global health issue, underscoring the necessity of timely diagnosis and treatment. The presumptive examination primarily relies on microscopic sputum smear tests, notably using Ziehl-Neelsen (ZN) staining. The existing methodologies of tuberculosis bacilli detection from bright-field microscopic sputum smear images suffer from low level of automation, inadequate segmentation performance and limited classification accuracy. In this paper, we propose an efficient methodology for tuberculosis bacilli identification that combines deep learning–based segmentation with a machine learning–based ensemble model for classification. The approach begins with an enhanced U-Net architecture incorporating attention blocks and residual connections to segment microscopic sputum smear images, enabling precise extraction of Regions of Interest (ROIs). The extracted ROIs are subsequently classified by an ensemble classifier which combines three efficient machine learning classifiers, Support Vector Machine (SVM), Random Forest and Extreme Gradient Boost (XGBoost) resulting in an accurate identification of bacilli within the image. For experimentation, a newly created dataset consisting of microscopic sputum smear images obtained from ZN-stained slides are utilized, in addition to the existing public datasets. Evaluation through qualitative and quantitative metrics reveals that the proposed model delivers significantly better segmentation performance, higher classification accuracy, and enhanced automation, distinguishing it from existing methodologies. • An efficient methodology for tuberculosis bacilli identification, combining deep learning for segmentation and a machine learning-based ensemble model for classification. • The proposed model demonstrates superior segmentation performance, enhanced classification accuracy, and improved degree of automation. • For experimentation, a newly created dataset consisting of microscopic sputum smear images obtained from ZN-stained slides are utilized.