Machine learning-based intelligent methods for determining total microbial colony quantities and classifying typical fungal morphotypes in cultured indoor air samples

Airborne microbial contamination is a health concern to building occupants. Traditional manual morphological classification and counting methods are time-consuming and subjective. They are inadequate for large-scale monitoring of indoor airborne microbes. In this study, using integrated machine learning technologies and computer vision algorithms, we trained the YOLOv8 object detector on an original dataset of 3,000 cultured sample images collected from indoor air of different building environments to establish two types of related prediction models: one for counting total microbial colonies and the other for classifying specific fungal morphotypes. The counting model achieved high prediction accuracy on the validation set, with both precision and recall values above 0.9. When evaluated on independent test sets, a strong correlation between predicted and manually counted colonies (r 2 >0.9) was observed. On a per-samples basis, the model counted the correct numbers in 47% to 60% of the samples. The single-morphotype classification models can classify and quantify target fungal groups with varying validation precision (0.5 to 0.8) and recall (0.3 to 0.8) for the following frequently detected taxa in indoor environments: Cladosporium sp. , Rhizopus nigricans, Mucor sp. , Penicillium sp. , Aspergillus versicolor, and Aspergillus flavus . On independent test sets, these classification models correctly predicted the exact colony counts in 25% to 95% of samples, depending on the target group. The results demonstrate a promising future for automatically screening and counting viable microbial colonies, both total colony counts and specific morphotypes, in cultured indoor environmental samples.