Abstract:
BACKGROUND: ‘‘TBDx’’ is an innovative smear microscopy system that automatically loads slides onto a microscope, focuses
and digitally captures images and then classifies smears as positive or negative using computerised algorithms.
OBJECTIVES: To determine the diagnostic accuracy of TBDx, using culture as the gold standard, and compare this to a
microscopist’s diagnostic performance.
METHODS: This study is nested within a cross-sectional study of tuberculosis suspects from South African gold mines. All
tuberculosis suspects had one sputum sample collected, which was decontaminated prior to smear microscopy, liquid
culture and organism identification. All slides were auramine-stained and then read by both a research microscopist and by
TBDx using fluorescence microscopes, classifying slides based on the WHO classification standard of 100 fields of view (FoV)
at 4006 magnification.
RESULTS: Of 981 specimens, 269 were culture positive for Mycobacterium tuberculosis (27.4%). TBDx had higher sensitivity
than the microscopist (75.8% versus 52.8%, respectively), but markedly lower specificity (43.5% versus 98.6%, respectively).
TBDx classified 520/981 smears (53.0%) as scanty positive. Hence, a proposed hybrid software/human approach that
combined TBDx examination of all smears with microscopist re-examination of TBDx scanty smears was explored by
replacing the ‘‘positive’’ result of slides with 1–9 AFB detected on TBDx with the microscopist’s original reading. Compared
to using the microscopist’s original results for all 981 slides, this hybrid approach resulted in equivalent specificity, a slight
reduction in sensitivity from 52.8% to 49.4% (difference of 3.3%; 95% confidence interval: 0.2%, 6.5%), and a reduction in the
number of slides to be read by the microscopist by 47.0%.
DISCUSSION: Compared to a research microscopist, the hybrid software/human approach had similar specificity and positive
predictive value, but sensitivity requires further improvement. Automated microscopy has the potential to substantially
reduce the number of slides read by microscopists.
