Automated Detection of Tuberculosis Bacilli Using Deep Neural Networks with Sputum Smear Images

Tuberculosis (TB) is a serious contagious disease caused by a bacterium called Mycobacterium tuberculosis (M. tuberculosis). Microscopic examination remains the primary laboratory diagnostic method for early detection of TB, but this method is time-consuming and inefficient due to the manual operation. To accelerate the process, automatic detection models are proposed. In this work, we have detected M. tuberculosis reliably from the sputum smear images using an improved YOLOv5 (DA-YOLO) algorithm. The backbone of YOLOv5 was modified by applying the coordinate attention block and self-attention block: the former encodes both channel relationships and long-range dependencies with precise positional information; the latter (transformer module) can fully rely on the self-attention mechanism to compute feature representations. More importantly, a DA-YOLO is proposed to achieve end-to-end detection and enumeration of Mycobacterium tuberculosis. The combination of the two attention mechanisms enables the backbone network to get rich information and robust features. A proposed model DAYOLO gives an overall accuracy of 87.6%. Experimental results show that our algorithm outperforms state-of-the-art methods, which will help computer-aided tuberculosis diagnosis faster and more accurately.