Classification between live and dead foodborne bacteria with hyperspectral microscope imagery and machine learning

Authors: Park, Bosoon; Shin, Tae-Sung; WANG, BIN - Oak Ridge Institute For Science And Education (ORISE); MCDONOGH, BARRY - Trutag Technologies; FONG, ALEXANDRE - Trutag Technologies

Submitted to: Journal of Microbiological Methods
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 5/11/2023
Publication Date: 5/13/2023
Citation: Park, B., Shin, T., Wang, B., Mcdonogh, B., Fong, A. 2023. Classification between live and dead foodborne bacteria with hyperspectral microscope imagery and machine learning. Journal of Microbiological Methods. https://doi.org/10.1016/j.mimet.2023.106739.
DOI: https://doi.org/10.1016/j.mimet.2023.106739

Interpretive Summary: Foodborne diseases cause serious public health problems. They account for an estimated 9 million sicknesses, 56,000 hospitalizations, and 1,300 deaths every year in the United States. Specifically, Salmonella, Escherichia coli O157, Listeria monocytogenes, and Campylobacter cause more than 2 million estimated foodborne illnesses, 31,000 hospitalizations, and 700 deaths, with an estimated negative impact on human health of up to $11 billion each year. Therefore, it is crucial to characterize the behavior of microorganisms in food and provide science-based data to support decision making in the prevention and mitigation of food safety issues. Moreover, quantitative assessment of bacteria viability is critical to reducing pathogenic ontamination in food and the environment. Although polymerase chain reaction (PCR) analysis is an accurate method to identify bacteria, PCR cannot differentiate between live and dead bacteria. Recently, hyperspectral microscope imaging (HMI) was demonstrated to be an emerging method for rapid label-free detection with advantages of non-destruction, easy-to-operate, and high throughput. In this research, an HMI method with machine learning algorithms was evaluated to differentiate live and dead foodborne bacteria.

Technical Abstract: Hyperspectral microscope imaging with deep learning methods accurately distinguished between live and dead foodborne bacteria using their spectral and morphological features. Three deep learning models, Fusion-Net I, II and III, were developed and evaluated to classify live and dead bacterial cells of six pathogenic strains including Escherichia coli (EC), Listeria innocua (LI), Staphylococcus aureus (SA), Salmonella Enteritidis (SE), Salmonella Heidelberg (SH), and Salmonella Typhimurium (ST) using morphological and spectral characteristics of live and dead bacterial cells. Each model with inputs of average spectra and 546 nm band images of the cells was distinct by using strain information as (1) Fusion-Net I didn't require strain information to identify cell viability, (2) Fusion-Net II included strain as another input for robust performance over different strains, and (3) Fusion-Net III identified bacterial strain as another output before robust identification of cell viability. The classification accuracies with Fusion-Net I model were 100% for LI, SE, and ST, 98.4% for EC, 96% for SA, and 92.9% for SH. However, Fusion-Net II and III models classified dead cells with consistent 100% accuracy over all six strains. In addition, Fusion-Net III model identified cell strains with 96.9% accuracy as a dual-task model, suggesting that live foodborne bacteria could be accurately identified by hyperspectral microscope imaging prior to causing foodborne outbreaks.