Hyperspectral imaging of common foodborne pathogens for rapid identification and differentiation

Abstract Hyperspectral imaging (HSI) provides both spatial and spectral information of a sample by combining imaging with spectroscopy. The objective of this study was to generate hyperspectral graphs of common foodborne pathogens and to develop and validate prediction models for the classification of these pathogens. Four strains of Cronobacter sakazakii, five strains of Salmonella spp., eight strains of Escherichia coli, and one strain each of Listeria monocytogenes and Staphylococcus aureus were used in the study. Principal component analysis and kNN (k‐nearest neighbor) classifier model were used for the classification of hyperspectra of various bacterial cells, which were then validated using the cross‐validation technique. Classification accuracy of various strains within genera including C. sakazakii, Salmonella spp., and E. coli, respectively, was 100%; except within C. sakazakii, strain BAA‐894, and E. coli, strains O26, O45, and O121 had 66.67% accuracy. When all strains were studied together (irrespective of their genus) for the classification, only C. sakazakii P1, E. coli O104, O111, and O145, S. Montevideo, and L. monocytogenes had 100% classification accuracy, whereas E. coli O45 and S. Tennessee were not classified (classification accuracy of 0%). Lauric arginate treatment of C. sakazakii BAA‐894, E. coli O157, S. Senftenberg, L. monocytogenes, and S. aureus significantly affected their hyperspectral signatures, and treated cells could be differentiated from the healthy, nontreated cells.


| INTRODUC TI ON
Hyperspectral imaging (HSI) is an emerging technology that has a great potential in rapid detection and identification of foodborne pathogens. The HSI was originally designed for remote sensing and now has found its application in various fields such as astronomy, agriculture, pharmaceutical, and medicine (Gowen, O'Donnell, Cullen, Downey, & Frias, 2007;Smith, 2012). The increased food safety and defense concerns due to foodborne pathogens have reinforced the need for rapid pathogen detection and identification methods in food. Although traditional methods are still used for the detection and identification of pathogens in food, these methods are cumbersome, labor-intensive, expensive, and can take from 4 to 7 days to give confirmatory results. Therefore, rapid detection methods for foodborne pathogens, at least at the presumptive level, are required for the functioning of a safe and fast food supply chain.
The HSI is a nondestructive method of analyzing and detecting a specimen and combines imaging with spectroscopy to acquire both spatial and spectral information of a specimen (such as bacterial cells or colonies) by using visible near-infrared spectra (400 to 1,000 nm; Dale et al., 2013;Gowen et al., 2007). The HSI utilizes optical characteristics of specimens for the identification; therefore, HSI works on the interactions between light (visible and/or near-infrared light) and the molecular structure of the specimen. During HSI, hyperspectral images of a sample are captured at various predefined wavelengths; and reflectance, transmittance, absorbance, or fluorescence (depending on the molecular and chemical structure of a sample) are measured over this wavelength range to generate hyperspectral graphs. These hyperspectral graphs, used to develop a library for different organisms, can then be used for the rapid identification of an unknown sample.
Rapid and early detection of pathogens in food is vital at every stage of processing, especially for "zero tolerance" pathogens such as Escherichia coli O157:H7 and Big Six Shiga toxin-producing E. coli (STEC; O26, O45, O103, O111, O121, and O145) in ground beef, and Listeria monocytogenes in ready-to-eat foods (Buchholz & Mascola, 2001;FSIS, 2012a;2012b). Another sensitive area where early and faster detection of pathogens is vital is the detection of Salmonella spp. and Cronobacter sakazakii in powdered infant formula manufactured for newborns and infants.
The previous research done utilizing HSI for the rapid identification and differentiation of various pathogens has been conducted on a macro-scale using bacterial colonies grown on a media agar, but best to our knowledge, studies at microscale (at the single bacterial cell level) are scarce in the literature (Eady & Park, 2016). Using HSI, Yoon et al., (2013a), Yoon et al. (2013b) were able to identify Big Six STEC grown on Rainbow agar as pure and mix cultures at accuracy levels from 80% to 100%. Fiore, Casale, and Aureli. (2008) were able to discriminate maize kernels infected with Aspergillus niger and A. flavus from uninfected kernels using HSI. The U.S. Department of Agriculture's Agriculture Research Service (ARS) developed and successfully tried (under commercial conditions) a HSI system to detect and separate fecal contaminated poultry carcasses from the good carcasses (FSIS, 2013). Factors like morphology of colonies grown on media agar and presence of food components on food surfaces, along with individual or clusters of bacterial cells, can interfere with HSI and affect the hyperspectra. Therefore, HSI of bacteria at the cellular level should be studied to exclude interfering factors in generating hyperspectra of targeted bacteria.
It is also very relevant and important to study whether inactive or injured bacterial cells can be differentiated from active and live bacterial cells using HSI (Anderson, Reynolds, Ringelberg, Edwards, & Floey, 2008). Lauric arginate (LAE) is a generally recognized as safe (GRAS) food-grade antimicrobial exhibiting activity against a wide range of pathogens (Saini, Miguel, Marsden, Getty, & Fung, 2013). The LAE has been studied for its use in food as an antimicrobial treatment and as a sanitizer for food contact surfaces (Becerril, Manso, Nerin, & Gómez-Lus, 2013;Saini, 2012). The LAE disrupts the plasma membrane of bacterial cells, which alters the metabolic processes and inhibits normal cellular cycles (Bakal & Diaz, 2005). Therefore, it is essential to study the effect of antimicrobials, such as LAE, on bacterial cells and subsequently on the HSI of these cells treated with antimicrobials. Therefore, the main objective of this research was to study whether hyperspectral graphs/signatures obtained from HSI of individual bacterial cells could be used for the rapid presumptive identification and differentiation of various strains of pathogenic bacteria. This included the development and validation of classification models for common foodborne pathogens. The effect of antimicrobial (LAE) treatment of various strains on hyperspectral signatures of these bacteria was also studied.

| Experimental design
To study whether HSI can be used for the rapid identification and differentiation of various foodborne pathogens, four strains of C. sakazakii, five strains of Salmonella spp., eight strains of E. coli, and one strain each of L. monocytogenes and S. aureus were used in the study. All strains were individually grown on tryptic soy agar (TSA; Difco, Becton Dickinson Company) for isolation, and hyperspectral images of these cells from the isolated colonies were captured. These images were then used to generate hyperspectral graphs of respective bacterial cells that were stored in a reference library. Principal component analysis (PCA) and kNN (k Near Neighbor) classifier model were used for the classification of hyperspectra of various pathogens, followed by the "cross-validation" technique. For LAE treatment, one strain from each genus was grown in TSB for 24 hr and then treated with 2,000 ppm LAE followed by HSI. Three replications were conducted for each strain for generating hyperspectral graphs; within each replication, HSI of various strains was conducted randomly.

| Culture propagation
The seventeen Gram-negative (four C. sakazakii, five Salmonella spp., and eight Shiga toxin-producing E. coli) and two Gram-positive (L. monocytogenes and S. aureus) foodborne pathogenic bacteria used in this study are presented in Table 1. All strains were selected on the basis of risk and involvement in foodborne disease outbreaks, or isolated from the environment and food processing facilities. All cultures were stored on TSA at 4°C and confirmed using API ® 20E (for C. sakazakii, Salmonella spp., and E. coli strains; Biomérieux), API ® Lister (for L. monocytogenes), and API ® Staph (for S. aureus).

| Bacterial cell samples preparation and immobilization
A loop (1 μl) from stock cultures stored on TSA was individually transferred into 10 ml TSB and incubated for 24 hr at 35°C for L. monocytogenes and E. coli, and 37°C for C. sakazakii, Salmonella spp., and S. aureus. Incubated strains were then individually streaked on TSA for isolation and incubated for 24 hr at respective temperatures. For each strain, one loop from an isolated colony was mixed in 1 ml filtered (0.2 μm) sterilized HPLC grade water (Fisher Scientific) in a microcentrifuge tube by vortexing for ~1 min. Ten μl of vortexed samples was individually transferred on clean and sanitized 1-mm glass slides (Fisherfinest ® , Fisher Scientific), and immobilized by air drying in a biosafety cabinet (SterilGARD ® II, The Baker Company) for ~5 min. These immobilized bacterial cells on glass slides were then used for the HSI analysis.
According to the Food Safety and Inspection Service (FSIS) of the U.S. Department of Agriculture, LAE can be used in various meat and poultry products up to 200 ppm by weight of the finished product (FSIS, 2013). Based on the previous research conducted on LAE treatment of veal carcasses in our laboratory, in which ~2,000 ppm LAE solution was used to achieve a final 200 ppm LAE by weight of beef cut being treated, 2,000 ppm LAE concentration was selected for this study. A loop from the stock culture of all bacteria on TSA was individually transferred into 10 ml of TSB and incubated for 24 hr at respective temperatures. For all strains, a loop from the inoculated TSB was then transferred to 10 ml TSB tubes and again incubated for 24 hr at respective temperatures. For each strain, after inoculation, 3 ml of inoculated TSB was mixed with 6 ml of noninoculated TSB and 1 ml of LAE (to achieve the final concentration of 2,000 ppm of LAE) and stored for 5 min at 4°C. After LAE treatment, 5 ml of D/E Neutralizing Broth was immediately added to the tubes (to neutralize the solution and arrest the antimicrobial effect of LAE). Tubes were then centrifuged at 4,960 × g at −4°C for 15 min, supernatant was discarded, and pellets were washed with 10 ml 0.1% peptone solution and centrifuged again. After centrifugation, supernatant was discarded, and one loop from the pellet in centrifuge tube was used for HSI analysis of each strain.

| Hyperspectral graphs generation
A CytoViva ® microscope system (CytoViva, Inc.) used for capturing hyperspectral images in the study is presented in Figure 1.

Environment for Visualizing Images (ENVI, Exelis Visual Information
Solutions) software version 4.4 was used for analyzing acquired hyperspectral images and generating hyperspectral graphs.
Hyperspectral images of individual bacterial cells on air-dried glass slides were acquired by focusing the microscope at 4,000 × magnification and ENVI settings at: 0.5 s exposure time, low grain, and low spatial and spectral resolutions (as recommended by CytoViva ® personal communication). Hyperspectral images were acquired using "push broom" (also known as "line scan") technique in which the glass slide supporting the dried culture on the motorized stage of microscope was moved across the field of view of the microscope during image acquisition. Using ENVI, three bacterial cells were selected from the acquired images as regions of interest (ROI). Average scattering values at respective wavelengths of these three ROI were used to generate hyperspectral graphs at wavelengths ranging from 400 to 1,000 nm (at wavelength interval of ~1.29 nm resulting in 465 wavelength bands). These graphs were then stored in a reference library for future use and comparison/classification.

| Principle component analysis, and classification and validation of hyperspectral graphs
Preprocessing is the first important step in statistical analyses and classification of a spectral data set; however, no general guidelines or rules exist for selecting a particular preprocessing technique for a specific type of a data set (Scott, James, & Ali, 2007). The type of preprocessing technique used for a particular data set should aim TA B L E 1 Gram-positive and Gram-negative bacteria used in the study where, X ij is normalized scattering value at a particular wavelength; X ij is actual scattering value at a particular wavelength; min (X j ) is minimum scattering value; max (X j ) is maximum scattering value.
Unsupervised PCA of normalized hyperspectral graphs was conducted to reduce the dimensionality and redundancy in the data set to the unclassified sample are calculated, and the unclassified sample is assigned to the group of the nearest k neighbors (Scott et al., 2007). In the kNN classification model, "k" is referred to as the number of close neighbors or samples to be considered for the grouping of the unclassified sample. The value of "k" can be any positive numeric value; however, values from 1 to 3 are commonly used (Scott et al., 2007;Yoon et al., 2013a;2013b). These classification models were then validated using the cross-validation technique.

| RE SULTS AND D ISCUSS I ON
An example of the image of bacterial cells as visible under the field of view of the microscope and acquired by the CytoViva ® microscope system is presented in Figure 2a. Using ENVI software, this view was further zoomed at two levels as presented in Figure 2b,c.
The appearance of bacterial cells at different zoom levels after the ROI was selected for acquiring hyperspectral graphs is presented in To study the variation of hyperspectral graphs/signatures within a genus, HSI of different strains of C. sakazakii, E. coli, and Salmonella  Figure 3a.
The graph indicates that C. sakazakii P1 had a different scattering pattern (shape of the curve) at wavelengths <615 nm compared with the other strains; however, at wavelengths >615 nm, the scattering patron of C. sakazakii P1 was similar to that of C. sakazakii E1.  Among E. coli serotypes, although all serotypes had a different intensity of scattering, they followed a similar pattern throughout the wavelength range ( Figure 6). Approximate grouping of various serotypes of E. coli can be visualized from the PCA plot presented in Figure 7. Except E. coli O26, O45, and O121, which had classification accuracy of 66.67%, classification accuracy of all other E. coli serotypes was 100% (Table 2). Escherichia coli O26, O45, and O121 were misclassified as E. coli O103, O121, and O45, respectively, at 33.33%.
To study whether HSI could be used to differentiate various  To study the efficacy of HSI to differentiate bacteria when different strains from different genera are analyzed together, PCA ( Figure 9) and kNN classification of hyperspectral graphs of various strains of C. sakazakii, E. coli and Salmonella spp., and L. monocytogenes and S. aureus was conducted. Cross-validation accuracy of the classification of different strains is presented in were the most poorly classified strains with an accuracy of 0%. No specific differentiation among the Gram-positive and Gram-negative strains used in this study can be made from the PCA data presented in Figure 9. However, by including more strains of L. monocytogenes, S. aureus, and other Gram-positive bacteria in future studies, the accuracy of HSI to differentiate Gram-positive bacteria from Gramnegative can be studied.  (4); Escherichia coli O26 (5), O45 (6), O103 (7), O104 (8), O111 (9), O121 (10), O145 (11), and O157 (12); Salmonella Agona BAA-707 (13), Montevideo (14), Senftenberg (15), Tennessee (16), and Typhimurium (17) and seven different layers over the core of spores (Montville & Mathews, 2005).
Currently, preassembled and preprogrammed HSI systems, such as CytoViva ® , are quite expensive and can easily cost over $125,000. This high initial investment cost is one of the major factors that restricts the research and application of HSI in the field of food microbiology. However, this initial cost can be reduced by mounting commercially available hyperspectral cameras on regular laboratory compound microscopes and using predeveloped imaging software (such as ENVI). At the current technology, this self-assembly process can reduce the initial investment cost to less than one-fifth of a preassembled HSI system. Once the HSI is installed, the cost of running HSI analysis is considerably low, which includes the cost of isolation media and regular microbiology laboratory tools (such as loops, glass slides, cover slips, and biosafety cabinet). In terms of analyzing time, once the bacterial colonies are isolated on an appropriate agar, the hyperspectral imaging acquiring and analyzing time is less than 15 min, which can be reduced further with more research and developing standard methods.

| CON CLUS IONS
In conclusion, overall classification accuracy of bacterial strains used in this study was acceptable when kNN classification model and cross-validation were applied within a specific genus. However, poor classification of various strains when compared together strongly suggests a need of further investigation, and use of different classification models and validation techniques, to eliminate the experimental and statistical factors responsible for poor classification. The HSI protocols should also be tested and validated for isolation and immobilization of bacterial cells from various food matrices. Although at the present state of the technology, HSI cannot replace the traditional microbiological identification methods, but it can be used effectively at presumptive levels for the identification of pathogens.

ACK N OWLED G M ENT
We acknowledge CytoViva ® , Inc., especially Dr. James Beach and Mr. Jamie Uertz, for their constant help and suggestions during the entire project that helped in developing the protocols used in the study. This is the contribution number 18-602-J from the Kanas Agricultural Experiment Station, Manhattan, KS.

CO N FLI C T S O F I NTE R E S T
All authors declare no conflicts of interest in this article.

E TH I C A L S TATEM ENT
This study does not involve any human or mammal testing.