Novel application of artificial intelligence based astronomical centroiding to small particle tracking

Quantitative assessment of aerosol particle size distribution is an important variable for evaluating barrier mask efficacy and mitigating airborne pathogen transmission. However, real world aerosol imaging is compromised by optical noise, such as light scatter, background interference and variable illumination, limiting the reliability of conventional testing approaches. Herein, we propose a novel solution to this issue centred on star centroiding, an astronomy-derived approach built on two deep learning models, MobileUNet and ELUNet. These models were originally trained on noisy images to detect celestial objects, which resemble aerosol particles. The models were applied without retraining to laser-illuminated aerosol images acquired during mask efficacy experiments. These images were generated by laser illumination of aerosol produced from a 10% NaCl-in-water solution and exhaled through a medical mannequin connected to an artificial lung, with the dataset comprising both constant and plume emission conditions. Across 41 images containing 7,123 manually annotated aerosol particles, MobileUNet detected 99% (model-to-manual ratio) of all manually annotated aerosol particles and achieved 98% for precision, recall, and F1-score, whereas ELUNet detected 97% of particles with precision, recall, and F1-score of 98%, 96%, and 97%, respectively. MobileUNet additionally reduced computation time by approximately four-fold compared to ELUNet. These findings demonstrate that astronomical centroiding models can be directly translated to aerosol analysis maintaining high performance in real life uncontrolled and high-noise environments. This interdisciplinary approach provides a robust framework for aerosol quantification and objective assessment of mask efficacy, with potential applications in pandemic preparedness and airborne infection control.