A New Artificial Intelligence-Based Method for Identifying Mycobacterium Tuberculosis in Ziehl-Neelsen Stain on...

Area of Science:

• Medical Diagnostics
• Computer Vision
• Pathology

Background:

• Accurate identification of Mycobacterium species is critical for diagnosing tuberculosis.
• Manual microscopic examination of Ziehl-Neelsen stained slides for mycobacteria is time-consuming and labor-intensive.
• Existing artificial intelligence (AI) methods for acid-fast bacilli (AFB) detection have limitations.

Purpose of the Study:

• To develop and validate an automated AI-based method for identifying mycobacteria in Ziehl-Neelsen stained slides.
• To improve the efficiency and accuracy of tuberculosis diagnosis.
• To assist pathologists by highlighting areas of interest on whole slide images (WSIs).

Main Methods:

• Development of a custom computer vision architecture utilizing extensive image augmentation techniques.
• Training and validation on a large dataset of over 260,000 positive and 700,000,000 negative patches from 510 whole slide images (WSIs).
• Internal validation on 286,000 patches and clinical testing on 60 ZN slides, comparing AI-assisted pathologist performance against manual review.

Main Results:

• The AI architecture achieved an area under the receiver operating characteristic curve (AUC) of 0.977.
• The AI-assisted method demonstrated 98.33% accuracy, 95.65% sensitivity, and 100% specificity in clinical testing.
• The AI-assisted approach outperformed previous AI-based methods for AFB detection.

Conclusions:

• The developed AI-based method significantly enhances the accuracy and efficiency of mycobacteria identification in pathology.
• This automated system shows great potential to support pathologists in diagnosing tuberculosis, improving patient outcomes.
• The AI tool provides a reliable and high-performing solution for acid-fast bacilli detection in ZN-stained slides.