Use of a convolutional neural network for direct detection of acid-fast bacilli from clinical specimens

Mycobacteria, including Mycobacterium tuberculosis (MTB) and non-tuberculosis mycobacteria (NTM), are important causes of infectious disease and cause significant mortality and morbidity globally. Fast detection is extremely important to reduce transmission and mortality associated with these infectious agents. Manual smear microscopy is a cost-effective tool for diagnosing and monitoring of these organisms; however, it is labor-intensive and requires highly-trained personnel. We present the development of an artificial intelligence computer vision process using a deep convolutional neural network to detect acid-fast bacilli (AFB) from Kinyoun acid-fast stained slides. We collected 231 clinical specimens between August 2023 and June 2024. Following acid-fast staining, whole slide images (WSI) were digitized, and AFB organisms were manually annotated. A machine learning computer vision model was trained using 11,411 annotated organisms across 109 WSI. Model predictions were correlated with final culture-confirmed results. The final model estimated AFB density per 1000 x microscope field of view (FOV). Using a density threshold of ≥10^-2 AFB/1000xFOV (corresponding to 1 + per Clinical and Laboratory Standards Institute (CLSI) guideline M48) to predict positive culture results, the model correctly classified 68% of validation slides, with a sensitivity of 79% and specificity of 63%. Manual AO compared to final culture read showed sensitivity of 76% and specificity of 96%. Although performance of our model was not sufficient to be clinically implemented in our laboratory, our study provides a framework for AI-based AFB detection and a publicly available data set to support future advancements in automated detection of AFB.IMPORTANCEWe present the development of an artificial intelligence model to detect acid-fast bacilli (AFB) directly from stained clinical smears. While the model's current performance requires further improvement to be clinically useful in our lab, we detail our approach and share our expertly annotated data set to support future advancements in this area. By building on our work, researchers can develop better algorithms to improve the diagnosis of AFB, reducing the burden on laboratory staff and improving diagnostic speed and accuracy of these medically important organisms.