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

Acid-fast bacillus (AFB) detection by microscopy remains a cornerstone of laboratory diagnosis for mycobacterial infections, despite the limitations of variable sensitivity, labor-intensive workflows, and interobserver variability. Recent advances in artificial intelligence, particularly deep learning, offer promising solutions to automate and augment AFB detection on microscopy images.

A systematic review was conducted in accordance with Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines, with protocol registration in PROSPERO (registration No. CRD42025633886). PubMed and Embase were searched for studies published in the past decade that applied deep learning-based methods to microscopy-based AFB detection. Eligible studies included original research and review articles using Ziehl-Neelsen-stained or auramine-stained specimens and reporting quantitative diagnostic performance metrics. Data on model architectures, annotation strategies, evaluation metrics, and accuracy were extracted. A forest plot was constructed to summarize pooled accuracy estimates across eligible studies.

Forty-nine studies met the inclusion criteria. Deep learning models, including convolutional neural networks, object detectors, and transformer-based architectures, demonstrated high accuracies (80%-98%), despite substantial heterogeneity. Weakly supervised learning, multiview and z-stack imaging, and domain adaptation techniques are being used to address variability in staining, focus, and laboratory conditions. Object detection and segmentation models have enabled improved bacillus localization and quantification, while foundation models and transfer learning approaches are reducing annotation burden and improving generalizability.

Deep learning-based AFB detection systems demonstrate high accuracy and strong potential to enhance tuberculosis microscopy. Heterogeneity in study design and limited external validation, however, highlight the need for standardized evaluation and multicenter clinical studies before widespread implementation.