Automated segmentation of foodborne bacteria from chicken rinse with hyperspectral microscope imaging and deep learning methods

Authors: Park, Bosoon; Shin, Tae-Sung; KANG, RUI - Jiangsu Academy Agricultural Sciences; FONG, ALEXANDRE - Trutag Technologies; MCDONOGH, BARRY - Trutag Technologies; Yoon, Seung-Chul

Submitted to: Computers and Electronics in Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 3/22/2023
Publication Date: 3/31/2023
Citation: Park, B., Shin, T., Kang, R., Fong, A., Mcdonogh, B., Yoon, S.C. 2023. Automated segmentation of foodborne bacteria from chicken rinse with hyperspectral microscope imaging and deep learning methods. Computers and Electronics in Agriculture. https://doi.org/10.1016/j.compag.2023.107802.
DOI: https://doi.org/10.1016/j.compag.2023.107802

Interpretive Summary: Foodborne illness has been a significant threat to public health. Since a leading cause of foodborne illness is food contamination with pathogenic bacteria, it is crucial to identify such bacteria in potentially contaminated food as early as possible in food supply chain. A traditional method for bacterial detection involves biochemical identification after culturing microorganisms in nutrient-enriched growth media that requires high cost for target-specific reagents and takes time for the result. Optical techniques such as optical coherence tomography and hyperspectral microscope imaging (HMI) have shown their potential to improve conventional methods using computer vision technologies. However, bacterial detection with HMI requires a task for segmenting single-cell bacteria to extract the spectral and spatial signatures of bacteria from hyperspectral image (hypercube). The segmentation for a hypercube involves identifying the pixels of single-cell bacteria in the image followed by creating the corresponding mask image with a binary image to distinguish bacteria from the background. Because a hypercube of bacteria contains lots of cells, manual processing to identify and segment single cells in a hypercube are time-consuming and inconsistent. Recently, HMI and deep learning (DL) have become a combination for unveiling food safety analysis. In this study, we developed DL methods for automated segmentation of single-cell bacteria in hyperspectral images by determining parameters for auto-segmentation through visual validation of hypercubes, creating reference mask images to identify single cells of foodborne bacteria efficiently.

Technical Abstract: Visible/near-infrared hyperspectral microscope imaging (HMI) has provided spectral-spatial features to identify pathogenic bacteria with high accuracy. But the bacterial detection with HMI requires accurate segmentation of single-cell bacteria from hyperspectral image (hypercube). In this study a robust technique was developed and evaluated to automatically segment single-cell pathogenic bacteria using deep learning and image processing. The proposed method consists of two steps as 1) bacterial segmentation with a deep learning model and 2) single-cell identification by ellipse fitting. Bacterial strains including Escherichia coli, Listeria, Salmonella, and Staphylococcus were prepared to obtain hyperspectral images of bacterial cells under different growth conditions with a Fabry–Perot Interferometer (FPI) HMI system. Based on the hypercube, four deep learning models including U-Net, residual U-Net (ResU-Net), attention gate residual U-Net (AGResU-Net), and attention-gated recurrent residual U-Net (AGR2U-Net) were employed for bacterial cell segmentation. AGR2U-Net with deblurred input images and -1 image padding performed better than other models with 94.1% mean intersection over union and visual inspection confirmed that segmented images with the model were identical to the ground-truth mask images. Also, ellipse fitting and goodness-of-fit evaluation rejected objects that should not be included in single-cell segmentation successfully. In addition, the robustness of the proposed method was confirmed because its segmentation accuracy and quality were moderately invariant with image blurriness and sample growth conditions. This accurate and robust auto-segmentation technique streamlined the detection of pathogenic bacteria with FPI-HMI by reducing processing time from raw hypercube acquisition to classification with 15 sec.