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

Area of Science:

• Medical Imaging
• Pathology
• Artificial Intelligence

Background:

• Tuberculosis (TB) poses a significant occupational hazard to autopsy staff, with incidence rates 3-5 times higher than for clinical personnel.
• South Korea faces a high TB burden, ranking highest in incidence and third in mortality among OECD countries in 2020.
• Traditional Ziehl-Neelsen staining for acid-fast bacilli (AFB) requires laborious, time-consuming microscopic examination at 1000× magnification.

Purpose of the Study:

• To develop and evaluate a computer-aided diagnosis (CAD) system for efficient TB detection in histopathological samples.
• To reduce the diagnostic time and improve the accuracy of TB screening in autopsy settings.
• To mitigate the occupational infection risk for autopsy staff and expedite cause-of-death determination.

Main Methods:

• Development of a CAD system utilizing nine neural networks trained on histopathological images.
• Training and evaluation performed using Ziehl-Neelsen stained slides at 400× magnification.
• Assessment of the model's ability to detect Mycobacterium tuberculosis (M. tuberculosis) in autopsy samples.

Main Results:

• The 'N' model demonstrated superior performance, achieving 99.77% accuracy per patch and 90% accuracy per slide.
• The CAD system effectively identified M. tuberculosis in evaluated slides.
• The AI model showed potential for preliminary TB screening, aiding pathologists.

Conclusions:

• The developed CAD system can significantly enhance the efficiency of TB diagnosis in autopsy work.
• This AI-driven approach aids pathologists in preliminary screening, reducing diagnostic workload and time.
• The research contributes to minimizing infection risks for autopsy staff and facilitating rapid death investigations.