Evaluation of an AI-Based TB AFB Smear Screening System for Laboratory Diagnosis on Routine Practice

Area of Science:

• Medical Diagnostics
• Artificial Intelligence in Medicine
• Microbiology

Background:

• Tuberculosis (TB) diagnosis relies on acid-fast bacilli (AFB) staining, which is labor-intensive and has low sensitivity.
• Artificial intelligence (AI) offers potential to enhance TB smear microscopy accuracy and efficiency.

Purpose of the Study:

• To evaluate the performance of an AI-powered automated system for detecting AFB in TB smear microscopy.
• To compare the AI system's diagnostic capabilities against manual screening methods.

Main Methods:

• An AI-based automated system, including a microscopic scanner and recognition program, was used to analyze 5930 TB smears.
• The system detected AFB and classified levels (0-4+), with image analysis per smear varying across stages.
• Performance metrics (accuracy, sensitivity, specificity) were assessed, including analysis after excluding poor-quality smears.

Main Results:

• Initial evaluation showed 91.3% accuracy, 60.0% sensitivity, and 95.7% specificity (120 images/smear).
• Performance improved to 93.7% accuracy, 77.4% sensitivity, and 96.6% specificity (200 images/smear).
• After quality control, accuracy reached 95.2%, sensitivity 85.7%, and specificity 96.9%, with 85 positive smears recovered from initial negative manual screening.

Conclusions:

• The automated TB system demonstrates higher sensitivity and laboratory efficiency compared to manual microscopy, especially with quality-controlled smear preparation.
• AI-driven automated TB smear screening can serve as an effective preliminary screening tool before manual microscopy.