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

Area of Science:

• Food Microbiology
• Spectroscopy
• Machine Learning

Background:

• Accurate identification of live foodborne bacteria is crucial for food safety.
• Current methods for bacterial viability assessment can be time-consuming.
• Novel techniques are needed for rapid and reliable detection.

Purpose of the Study:

• To develop and evaluate deep learning models for distinguishing live from dead foodborne bacteria using hyperspectral imaging.
• To assess the models' performance across six pathogenic bacterial strains.
• To explore the potential of these models for dual-task applications in food safety.

Main Methods:

• Hyperspectral microscope imaging was employed to capture spectral and morphological features of bacterial cells.
• Three deep learning models (Fusion-Net I, II, III) were developed using average spectra and 546 nm band images.
• The models were trained and tested on six pathogenic bacterial strains: Escherichia coli, Listeria innocua, Staphylococcus aureus, Salmonella Enteritidis, Salmonella Heidelberg, and Salmonella Typhimurium.

Main Results:

• Fusion-Net I achieved high accuracy (100% for some strains) in identifying live bacterial cells.
• Fusion-Net II and III demonstrated robust 100% accuracy in classifying dead cells across all strains.
• Fusion-Net III successfully identified bacterial strains with 96.9% accuracy, indicating dual-task capability.

Conclusions:

• Hyperspectral imaging coupled with deep learning provides an effective tool for bacterial viability assessment.
• This approach offers a faster and more reliable alternative for food safety inspections.
• The developed models show promise for early detection of foodborne pathogens and prevention of outbreaks.