Classification of <i>Mycobacterium tuberculosis</i> in Images of ZN-Stained Sputum Smears

Screening for tuberculosis (TB) in low- and middle-income countries is centered on the microscope. We present methods for the automated identification of Mycobacterium tuberculosis in images of Ziehl-Neelsen (ZN) stained sputum smears obtained using a bright-field microscope. We segment candidate bacillus objects using a combination of two-class pixel classifiers. The algorithm produces results that agree well with manual segmentations, as judged by the Hausdorff distance and the modified Williams index. The extraction of geometric-transformation-invariant features and optimization of the feature set by feature subset selection and Fisher transformation follow. Finally, different two-class object classifiers are compared. The sensitivity and specificity of all tested classifiers is above 95% for the identification of bacillus objects represented by Fisher-transformed features. Our results may be used to reduce technician involvement in screening for TB, and would be particularly useful in laboratories in countries with a high burden of TB, where, typically, ZN rather than auramine staining of sputum smears is the method of choice.