Skip to main content
ARS Home » Southeast Area » Athens, Georgia » U.S. National Poultry Research Center » Quality & Safety Assessment Research » Research » Publications at this Location » Publication #358248

Research Project: Develop Rapid Optical Detection Methods for Food Hazards

Location: Quality & Safety Assessment Research

Title: Identifying non-O157 Shiga toxin-producing Escherichia coli (STEC) using deep learning methods with hyperspectral microscope images

Author
item KANG, RUI - US Department Of Agriculture (USDA)
item Park, Bosoon
item CHEN, KUN-JIE - Nanjing Agricultural University

Submitted to: Spectrochimica Acta
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 7/13/2019
Publication Date: 7/16/2019
Citation: Kang, R., Park, B., Chen, K. 2019. Identifying non-O157 Shiga toxin-producing Escherichia coli (STEC) using deep learning methods with hyperspectral microscope images. Spectrochimica Acta. https://doi.org/10.1016/j.saa.2019.117386.
DOI: https://doi.org/10.1016/j.saa.2019.117386

Interpretive Summary: A series of foodborne pathogen outbreaks increase threat to the public, and major outbreaks of foodborne illness are caused by several pathogens includes Salmonella spp., Campylobacter spp. and Shiga toxin-producing Escherichia coli (STEC). Although conventional bacterial detection methods such as aerobic plate count (APC) and standard plate count (SPC) are accurate and still the “gold standard” in preventing further threats to the public, these methods are laborious and time consuming. Here, we applied deep learning method called “Auto-Encoder” to identify pathogenic foodborne bacteria with hyperspectral microscope images at the cellular level as a big data analysis tool. Since hyperspectral image data from foodborne bacteria are too big to analyze with conventional statistical methods, an innovative method using machine learning algorithms has been employed for large-scale data processing and analysis. During the process of deep auto-encoder (DAE) model, the scattered intensity values of bacteria were transformed for classification models. In comparison of conventional statistical data analysis methods, deep learning methods proposed in this study improved the classification accuracy up to 95%. Thus, a deep learning method has the potential to analyze big hyperspectral microscopic image data for STEC serogroup classification.

Technical Abstract: Hyperspectral microscope imaging (HMI) method, which provides both spatial and spectral information of Non-O157 Shiga toxin-producing Escherichia coli (STEC) bacterial cells, was presented as an efficient tool to classify serogroups at the cellular level. Using hyperspectral microscope image data collected by an electron multiplied charge coupled device (EMCCD) camera with acousto-optic tunable filters platform, deep learning methods were evaluated to analyze big data from “Big Six” STEC bacteria. The spectral information extracted from ground truth regions of interest (ROI) were compiled for various foodborne pathogen samples. The aim of this research was to classify STEC serogroups using a deep neural network (DNN) that consisted of deep auto-encoder (DAE) and Softmax regression (SR). The DAE performed as a non-linear principal component analysis (PCA) to extract principal features from original data using layer-wise training to acquire deeper features in unsupervised way. Consequently, the SR was employed to classify the features obtained from deeper learning process. A back propagation fine-tuning algorithm was used to improve model accuracy, resulting in classification accuracy of 95.7% which was higher than 89.3% for linear discriminant analysis (LDA) and 83.1% for support vector machine (SVM) methods.