A Study of Mycobacterium tuberculosis Detection Using Different Neural Networks in Autopsy Specimens

Abstract

Tuberculosis (TB) presents a substantial health risk to autopsy staff, given its three to five times higher incidence of TB compared to clinical staff. This risk is notably accentuated in South Korea, which reported the highest TB incidence rate and the third highest TB mortality rate among OECD member countries in 2020. The standard TB diagnostic method, histopathological examination of sputum or tissue for acid-fast bacilli (AFB) using Ziehl–Neelsen staining, demands microscopic examination of slides at 1000× magnification, which is labor-intensive and time-consuming. This article proposes a computer-aided diagnosis (CAD) system designed to enhance the efficiency of TB diagnosis at magnification less than 1000×. By training nine neural networks with images taken from 30 training slides and 10 evaluation slides at 400× magnification, we evaluated their ability to detect M. tuberculosis. The N model achieved the highest accuracy, with 99.77% per patch and 90% per slide. We discovered that the model could aid pathologists in preliminary TB screening, thereby reducing diagnostic time. We anticipate that this research will contribute to minimizing autopsy staff’s infection risk and rapidly determining the cause of death.