Seung-Chul Yoon

Hyperspectral Reflectance Imaging for Detecting a Foodborne Pathogen: Campylobacter

This article is concerned with the development of a hyperspectral reflectance imaging technique for detecting and identifying one of the most common foodborne pathogens, Campylobacter. Direct plating using agars is an effective tool for laboratory tests and analyses of microorganisms. The morphology (size, growth pattern, color, etc.) of colonies grown on agar plates has been widely used to tentatively differentiate organisms. However, it is sometimes difficult to differentiate target organisms like Campylobacters from other contaminants grown together on the same agar plates. A hyperspectral reflectance imaging system operating at the visible and near-infrared (VNIR) spectral region from 400 nm to 900 nm was set up to measure spectral signatures of 17 different Campylobacter and non-Campylobacter subspecies. Protocols for culturing, imaging samples and for calibrating measured data were developed. The VNIR spectral library of all 17 organisms commonly encountered in poultry was established from calibrated hyperspectral reflectance images. A pattern classification algorithm was developed to locate and identify 48 h cultures of Campylobacter and non-Campylobacter contaminants on background agars (blood agar and Campy-Cefex) with over 99% accuracy. The Bhattacharyya distance, a statistical separability measure, was used to predict the performance of the pattern classification algorithm at a few wavelength bands chosen by the principal component analysis (PCA) band weightings. This research has a potential to be expanded to detect other pathogens grown on agar media.

Detection by Hyperspectral Imaging of Shiga Toxin–Producing Escherichia coli Serogroups O26, O45, O103, O111, O121, and O145 on Rainbow Agar

The U.S. Department of Agriculture, Food Safety Inspection Service has determined that six non-O157 Shiga toxin-producing Escherichia coli (STEC) serogroups (O26, O45, O103, O111, O121, and O145) are adulterants in raw beef. Isolate and phenotypic discrimination of non-O157 STEC is problematic due to the lack of suitable agar media. The lack of distinct phenotypic color variation among non-O157serogroups cultured on chromogenic agar poses a challenge in selecting colonies for confirmation. In this study, visible and near-infrared hyperspectral imaging and chemometrics were used to detect and classify non-O157 STEC serogroups grown on Rainbow agar O157. The method was first developed by building spectral libraries for each serogroup obtained from ground-truth regions of interest representing the true identity of each pixel and thus each pure culture colony in the hyperspectral agar-plate image. The spectral library for the pure-culture non-O157 STEC consisted of 2,171 colonies, with spectra derived from 124,347 of pixels. The classification models for each serogroup were developed with a k nearest-neighbor classifier. The overall classification training accuracy at the colony level was 99%. The classifier was validated with ground beef enrichments artificially inoculated with 10, 50, and 100 CFU/ml STEC. The validation ground-truth regions of interest of the STEC target colonies consisted of 606 colonies, with 3,030 pixels of spectra. The overall classification accuracy was 98%. The average specificity of the method was 98% due to the low false-positive rate of 1.2%. The sensitivity ranged from 78 to 100% due to the false-negative rates of 22, 7, and 8% for O145, O45, and O26, respectively. This study showed the potential of visible and near-infrared hyperspectral imaging for detecting and classifying colonies of the six non-O157 STEC serogroups. The technique needs to be validated with bacterial cultures directly extracted from meat products and positive identification of colonies by using confirmatory tests such as latex agglutination tests or PCR.

Fisher Linear Discriminant Analysis for Improving Fecal Detection Accuracy with Hyperspectral Images

Detection of fecal contamination in the visceral cavity of broiler carcasses is important for food safety to protect consumers from food pathogens. The simple ratio of the reflectance values of a 565 nm image to a 517 nm image was effective for identifying cecal contaminants in the visceral cavity. Since the accuracy of detection algorithms for identifying cecal contaminants varies with fecal threshold values, determination of the optimum threshold is crucial for detecting fecal contaminants during poultry processing. The dynamic threshold method using Fisher linear discriminant analysis (FLDA), along with the simple multispectral image ratio with Gaussian window averaging (10 nm FWHM bandwidth), performed better (98.9% detection accuracy with 1.1% omission error) than the static threshold method to identify cecal contaminants. The static threshold method with a threshold value close to the average of the estimated dynamic threshold values achieved 97.3% detection accuracy with 2.7% omission error. Because of uncertainty of fecal threshold and the trade-off between missed contaminants and false positives, the dynamic thresholding method using FLDA was useful for cecal contaminant detection. In addition, FLDA can be implemented to determine and update fecal threshold values for on-line inspection at poultry processing plants.

Hyperspectral Imaging for Differentiating Colonies of Non-0157 Shiga-Toxin Producing Escherichia Coli (STEC) Serogroups on Spread Plates of Pure Cultures

Direct plating on solid agar media has been widely used in microbiology laboratories for presumptive-positive pathogen detection, although it is often subjective and labour-intensive. Rainbow agar is a selective and differential chromogenic medium used to isolate pathogenic Escherichia coli (E. coli) strains. However, it is challenging to differentiate colonies of the six representative pathogenic non-0157 Shiga-toxin producing E. coli (STEC) serogroups (026, 045, 0103, 0111, 0121 and 0145) on Rainbow agar due to the phenotypic differences and the presence of background microflora. Therefore, there is a need for a method or technology to objectively, rapidly and accurately perform high-throughput screening of non-0157 STEC colonies on agar plates. In this paper, we report the development of a visible-near infrared hyperspectral imaging technique and prediction model for differentiating colony types of the six non-0157 STEC serogroups in spread plates of pure cultures inoculated on Rainbow agar. The prediction model was based on supervised linear classification of factor scores obtained by principal component analysis (PCA). Both PCA-MD (Mahalanobis distance) and PCA-kNN (k-nearest neighbour) classifiers were used for model development. From the 24 hyperspectral images measured from two replicates, 51,173 spectral samples were collected from 1421 colonies. Chemometric preprocessing methods and other operating parameters, such as scatter correction, first derivative, moving average, sample size and number of principal components (PCs), were compared with a classification and regression tree (CART) method, configured as a classification tree and followed by brute-force searching from candidates selected by the CART. The number of PCs, scatter correction and moving average were selected as the most important elements to consider before selecting a set of candidate models. Cross-validation (CV), such as hold out and k-fold CV, was used to validate the prediction performance of candidate models. Serogroups 0111 and 0121 consistently showed over 99% classification accuracy regardless of the classification algorithms. However, the classification accuracies of serogroups 026, 045, 0103 and 0145 showed varying results from 84% up to 100%, depending on which preprocessing treatment and prediction model were adopted. The best overall mean classification accuracy of 95.06% was achieved with PCA-kNN (k=3), six PCs, five-pixel sample size defined around each colony centre, standard normal variate and detrending, first derivative with 11-point gaps and moving average with 11-point gaps. Future studies will focus on automating colony segmentation, further improving detection accuracy of the developed models, expanding the spectral library to include background microflora from ground beef and developing prediction models to detect the target bacteria in the presence of background microflora.

Modified pressure imaging for egg crack detection and resulting egg quality

Cracks in the shell surface compromise the primary barrier for external microbial contamination of the egg. Microcracks are very small cracks in the shell surface that are difficult to detect by human graders. New technology has been developed that uses modified pressure and imaging to detect microcracks in eggs. Research has shown the system to have an accuracy of 99.6% in detecting both cracked and intact eggs. A study was undertaken to determine if quality differences existed between modified pressure imaged and control eggs during extended cold storage. Three replicates were conducted with eggs stored at 4 degrees C for 5 wk with weekly quality testing. The physical quality factors monitored were Haugh units, albumen height, egg weight, shell strength, vitelline membrane strength and elasticity, and whole egg total solids. All measurements were conducted on individual eggs (12/treatments per replicate) each week with the exception of whole egg solids, which were determined from 3 pools (4 eggs each)/treatment per replicate each week. Percentage of whole egg total solids was the only significant difference (P < 0.05) between treatments (23.65% modified pressure imaged and 23.47% control). There was a significant difference (P < 0.05) for egg weight between replicates (60.82, 58.02, and 60.58 g for replicates 1, 2, and 3, respectively). Therefore, imaging eggs in the modified pressure system for microcrack detection did not alter egg quality during extended cold storage. Utilizing the modified pressure crack detection technology would result in fewer cracked eggs reaching the consumer, consequently enhancing food safety without affecting product quality.

Optimization of Fecal Detection Using Hyperspectral Imaging and Kernel Density Estimation

This article addresses the development of an iterative search algorithm to find an optimal threshold to detect surface contaminants on poultry carcasses for a real-time multispectral imaging application. Previous studies showed that a band-ratio algorithm with a 517 nm band and a 565 nm band could detect contaminants on the surface of poultry carcasses. In this study, thresholding for the band-ratio algorithm was optimized in a statistical sense. A fundamental problem of the thresholding is that there is a theoretical performance bound from the standpoint of statistical hypothesis testing. In a Neyman-Pearson (NP) framework, a lower bound of detection accuracy can be determined for minimizing false positives. An iterative search algorithm was designed to obtain an optimal threshold in the NP framework. For the design of the search algorithm, statistical density distributions of fecal and non-fecal image data were estimated by kernel density estimation, and characterized by edge models on a projection axis perpendicular to a linear decision boundary. Three necessary criteria were investigated for the selection of the optimum threshold of the band-ratio algorithm. Numerical simulations with hyperspectral poultry images showed that the optimum threshold was 1.05.

Modified Pressure System for Imaging Egg Cracks

One aspect of grading table eggs is shell checks or cracks. Currently, USDA voluntary regulations require that humans grade a representative sample of all eggs processed. However, as processing plants and packing facilities continue to increase their volume and throughput, human graders are having difficulty matching the pace of the machines. Additionally, some plants also have a problem with micro-cracks that the graders often miss because they are very small and hard to see immediately post-processing but grow and become readily apparent before they reach market. An imaging system was developed to help the grader detect these small micro-cracks. The imaging system utilized one image captured at atmospheric pressure and a second at a slight negative pressure to enhance the crack and make detection much easier. A simple image processing algorithm was then applied to the ratio of these two images, and the resulting image, containing both cracked and/or intact eggs were color-coded to simplify identification. The imaging system was capable of imaging 15 eggs in about 0.75 s, and the algorithm processing took about another 10 s. These times could easily be reduced with a compiled program specifically written for the application. In analyzing 1000 eggs, the system was 99.6% accurate overall with only 0.3% false positives, compared to 94.2% accurate overall for the human graders with 1.2% false positives. An international patent on the system has been filed, and further automation of the system is needed.

Acousto-Optic Tunable Filter Hyperspectral Microscope Imaging Method for Characterizing Spectra from Foodborne Pathogens

A hyperspectral microscope imaging (HMI) method, which provides both spatial and spectral information of bacterial cells, was developed for foodborne pathogen detection. The AOTF-based hyperspectral microscope imaging system can be effective for characterizing spectral properties of biofilms formed by Salmonella enteritidis as well as Escherichia coli. The intensity of spectral images and the pattern of intensity distribution of spectra varied with system parameters (integration time as well as gain) of the HMI system. Preliminary results demonstrated that determination of optimum parameter values of the HMI system and the integration time, which must be no more than 250 ms, are important for quality image acquisition from biofilms formed by S. enteritidis. Among the 89 contiguous spectral images between 450 and 800 nm, the intensity of images at 458, 494, 522, 550, 574, 590, and 670 nm were distinctive for biofilm of S. enteritidis, whereas the intensity of spectral images at 546 nm was distinctive for E. coli with dark-field illumination with a metal halide light source. For more accurate comparison of intensity from spectral images, a calibration protocol of quantitative intensity comparison needs to be developed to standardize image acquisition using neutral-density filters and multiple exposures. For the identification or classification of unknown foodborne pathogen samples, ground truth region-of-interest pixels need to be selected for spectrally pure fingerprints from various foodborne pathogens such as E. coli and Salmonella species.

Bone Fragment Detection in Chicken Breast Fillets Using Transmittance Image Enhancement

This article is concerned with the detection of bone fragments embedded in de-boned skinless chicken breast fillets by modeling optical images generated by backlighting. Imaging of chicken fillets is often dominated by multiple scattering properties of the fillets. Thus, resulting images from multiple scattering are diffused, scattered, and low contrast. In this study, a combination of transmittance and reflectance hyperspectral imaging, which is a non-ionized and non-destructive imaging modality, was investigated as an alternative method to conventional transmittance x-ray imaging, which is an ionizing imaging modality. As a way of reducing the influence of light scattering on images and thus increasing the image contrast, the use of a structured line light was examined along with an image formation model that separated undesirable lighting effects from an image. The image formation model, based on an illumination-transmittance model, was applied for correcting non-uniform illumination effects so that embedded bones were more easily detected by a single threshold. An automated image processing algorithm to detect bones was also proposed. Experimental results with chicken breast fillets and bone fragments are provided. The detection accuracy of the developed technology was 100%. The false-positive rate was 10%.

Statistical Model-Based Thresholding of Multispectral Images for Contaminant Detection on Poultry Carcasses

Developing an algorithm to decide the presence or absence of fecal contamination on the surface of poultry carcasses is critical to food safety. The global threshold strategy for a band-ratio algorithm has been known to be limited to pixel-basis detection. In an attempt to develop a statistical decision rule for carcass-basis detection from multispectral images, probability density functions of both contaminated and uncontaminated materials were estimated by parametric and non-parametric methods. We found that uncontaminated poultry carcasses could be modeled by a Gaussian distribution, whereas contaminated materials were non-Gaussian. A kernel density estimator was used to analyze the non-Gaussian characteristic of the contaminated materials on a transformed projection axis. A linear mixture of the density functions was introduced to model the observations made on the projection axis. A new detection algorithm was designed using the mixture model and tested for 496 birds (248 dirty and 248 clean birds). A test on the sample birds revealed that the algorithm needed at least 12 contaminant pixels to reach the perfect detection results. The test also showed a false-positive rate of less than 5%.