Deep learning-based detection of acid-fast bacilli in microscopy images: a systematic review

Area of Science:

• Medical Diagnostics
• Artificial Intelligence in Medicine
• Microbiology

Background:

• Microscopy is crucial for diagnosing mycobacterial infections, but faces challenges like low sensitivity and high variability.
• Artificial intelligence (AI), specifically deep learning, presents an opportunity to automate and improve acid-fast bacillus (AFB) detection.

Purpose of the Study:

• To systematically review the application of deep learning methods for AFB detection in microscopy images.
• To assess the diagnostic performance and identify trends in AI-based AFB detection.

Main Methods:

• A systematic review of studies from the past decade was performed using PubMed and Embase.
• Studies focused on deep learning for Ziehl-Neelsen or auramine-stained AFB detection, extracting performance metrics.
• Data synthesis included model architectures, annotation strategies, and pooled accuracy estimation.

Main Results:

• Forty-nine studies were included, showing deep learning models achieved high accuracies (80%-98%).
• Various deep learning architectures (CNNs, object detectors, transformers) were employed.
• Techniques like weakly supervised learning and domain adaptation addressed image variability, while foundation models reduced annotation needs.

Conclusions:

• Deep learning-based AFB detection shows high accuracy and potential to enhance tuberculosis microscopy.
• Challenges include heterogeneity in study design and limited external validation.
• Standardized evaluation and multicenter clinical studies are necessary before widespread adoption.