Progress 10/01/23 to 09/30/24
Outputs
PROGRESS REPORT Objectives (from AD-416): 1. Develop imaging technologies to detect and identify plastics during poultry processing with hyperspectral imaging and artificial intelligence. 1A. Develop hyperspectral imaging technology for detection and identification of plastic foreign objects during poultry processing. 1B. Develop AI technology for enhanced detection and smart robotic removal of foreign materials in hyperspectral imagery during poultry processing. 2. Detection and identification of foodborne bacteria and toxins in poultry products with high-throughput hyperspectral microscopy and surface plasmon resonance imaging. 2A. Rapid monitoring of indicator microorganisms in poultry processing. 2B. Develop advanced hyperspectral microscope imaging (HMI) methods and system for label-free detection and identification of pathogens at the cellular level with no enrichment. 2C. Develop high-sensitive and selective immunoassay method and system for foodborne bacteria and toxin detection with surface plasmon resonance imaging. 3. Eliminate the production of semicarbazide in non-nitrofurazone treated poultry by optimization of antimicrobial treatments and/or alternative antimicrobials during processing. 4. Develop safe and effective poultry processing strategies (scalding- picking-evisceration procedures) to reduce foodborne contaminants (pathogens/chemical) and enhance the sustainability of poultry processing. 4a. Develop sustainable poultry processing using artificial intelligence (AI) technology to improve poultry food safety. 4b. Develop Internet of Things (IoT) technology with various sensing platforms and data analytics for smart poultry processing and safety. Approach (from AD-416): Research on poultry safety will focus on: 1) developing and validating early, rapid, sensitive, and/or high-throughput optical sensing techniques for detecting physical and biological hazards in poultry products, and 2) eliminating semicarbazide in non-nitrofurazone treated poultry by optimization of antimicrobial treatments. In research on optical detection of physical hazards, spectroscopic and hyperspectral imaging (HSI) technologies will be developed for detection and identification of plastic foreign objects. A robot rejector and control software will be developed to eliminate foreign materials (FM) when detected by HSI. Artificial intelligence (AI) technology will be developed for enhanced detection and smart robotic removal of FM during poultry processing through the development and evaluation of customized deep learning algorithms based on hyperspectral imaging. A vision-guided smart robotic manipulator will be designed and built to remove FM by self- learning AI algorithms. To develop methods and techniques for detecting and identifying biological hazards, time-lapse image data on pure-culture indicator organisms and poultry carcass rinses from different processing locations will be collected to build a library, which will be used for on- line counting of microcolonies to build prototype systems. To detect foodborne pathogens, hyperspectral microscope imaging (HMI) methods will be developed with a spectral library of various pathogens using two HMI platforms of acousto-optical tunable filter (AOTF) and Fabry-Perot interferometer (FPI). In accordance with optimization of parameters on HMI and hypercubes, a transportable HMI system will be developed embedded with AI-based software for classification and identification. To identify foodborne bacteria and toxins, a highly-sensitive and selective immunoassay method and system will be developed using surface plasmon resonance imaging (SPRi). Microfluidic devices will be designed, simulated and fabricated for bacterial enrichment and separation. Both materials and parameters to develop a 3D printed biosensor for multiplex detection of pathogenic bacteria and toxins will be optimized and evaluated with food samples. Finally, a portable 3D printing platform for biosensor fabrication by integrating sample enrichment cartridge, biochip and SPRi detector will be developed. To develop techniques for eliminating the production of semicarbazide (SEM) in non-nitrofurazone treated poultry, a methodology for SEM analysis in chicken meat and a data library relating poultry processing conditions to SEM formation will be developed. Specifically, SEM in chicken leg quarters obtained from multiple processing facilities will be analyzed and methods to eliminate SEM production in poultry products under processing conditions will be developed. Research was conducted to develop hyperspectral imaging technology for detection and identification of plastic foreign materials (FM) during poultry processing (Sub-objective 1A). Researchers made significant progress on an engineering project to develop a novel sensor fusion methodology that integrates deep learning and two powerful imaging techniques - color and hyperspectral imaging. The developed fusion technology utilizes deep learning algorithms based on You Only Look Once (YOLO) and Convolutional Neural Network (CNN) models. The YOLO version 8 Extra Large (YOLOv8x) and YOLOv7x were applied to color images, demonstrating good performance at detecting plastic pieces with 99% accuracy for 5 mm x 5 mm FM pieces and 80% accuracy for 2 mm x 2 mm FM pieces on chicken breast fillets. A one-dimensional CNN model was applied to analyze the detected FM objects in hyperspectral images (600 - 1700 nm) and predict their plastic types, providing a detection accuracy of 97% in identifying 12 different plastic types. This research underscores the importance of cutting-edge artificial intelligence (AI) technologies in tackling FM detection within the poultry industry. This sensor fusion methodology, with its proven effectiveness, will be further refined to enhance its impact, potentially increasing food safety standards and quality control measures for the poultry industry. Research was conducted to develop AI technology for enhanced detection and smart robotic removal of foreign materials (FM) in hyperspectral imagery during poultry processing (Sub-objective 1B). Researchers made progress on developing a high-performance deep learning (DL) model specifically designed for high-speed and high-throughput detection of FM in poultry meat. While DL models offer exceptional accuracy, their computational demands during inferencing time can hinder real-time predictions when dealing with the large volume of data generated by real- time hyperspectral imaging systems, especially for inline poultry processing applications. To address this challenge, researchers developed a real-time inferencing model based on NVIDIA TensorRT, a software development kit for high-performance DL inference on NVIDIA Graphics Processing Unit (GPU). The model was optimized for real-time processing with fixed numerical precisions on TensorRT. The optimized model for GPUs achieved excellent results for inference time performance, achieving 17 times faster processing compared to CPUs and 6 times faster than standard GPUs, without compromising detection accuracy. This development holds promise for further enhancing performance of real-time FM detection and classification using AI-based hyperspectral imaging in poultry processing. Progress was made on research for identification of the root cause of the generalization issue (Sub-objective 2B). The use of hyperspectral microscope imaging (HMI) with deep learning (DL) has proven effective in identifying pathogenic bacterial cells. However, models trained on HMI datasets from different years showed spectral-spatial bias, affecting their generalizability. Researchers investigated this issue and found that the bias was due to differences in brightness and blurriness of bacterial cells in hyperspectral images. To address this, researchers developed a software package to measure and maintain consistent brightness and blurriness during hyperspectral image acquisition. Implementing this software is expected to significantly reduce bias across datasets. Significant progress was made on research for development of a trustworthy detection model with uncertainty measurements (Sub-objective 2B). Like most AI models, the previous detection model showed low accuracy with adversarial data points despite high performance with data similar to the training set. To address this inconsistency, researchers (a) explored various methods to measure uncertainties in deep learning models, (b) implemented deep k-nearest neighbors using inductive conformal prediction as the optimal framework, and (c) developed a new bacterial detection model that leverages uncertainty measurements from sub-models with different modalities. This advanced model improved accuracy on adversarial data points (unseen data) from 40% to 89% while maintaining high accuracy on data similar to the training set (seen data). These findings suggest that (a) a reliable and generalizable detection model for foodborne bacteria can be constructed from non-generalizable sub-models using uncertainty measurements, and (b) adding more sub-models for additional modalities may further enhance performance on both seen and unseen data. Progress was made on research for development of the two-photon lithography (TPL) technique for the 3D printing of the desired microfluidic devices (Sub-objective 2C). TPL is a high-resolution 3D printing technique that uses high-intensity focused lasers to produce features smaller than the size of the focused light spot. Researchers developed the Projection TPL (P-TPL) technique, which is a high- throughput variant of TPL and demonstrated that leak-free interfaces can be fabricated, small-scale devices can be fabricated with TPL and then integrated with a larger chip, and to enable the printing of dense 3D structures using P-TPL. Microfluidic chips with sub-10 µm channels were printed using the TPL process, which was performed with a commercial Nanoscribe 3D printer. Printing was performed with a proprietary photoresist material that contains multifunctional acrylate oligomers. The printed material is optically transparent and enables imaging of the microchannels that are contained within these chips. Research to develop safe and effective poultry processing strategies (scalding-picking-evisceration procedures) to reduce foodborne contaminants (pathogens/chemical) and enhance the sustainability of poultry processing (Objective 4) was conducted. As part of a formal ARS agreement, university collaborators conducted trials on: 1) re-use of poultry processing wastewater for hydroponic plant production, 2) alternative uses of solid residues from poultry processing plants, 3) effects of broiler lairage and controlled atmosphere stunning (CAS) on poultry meat, 4) effects of CAS on metabolic function of broilers, 5) carcass quality at various points during processing, 6) effects of CAS on sensory attributes of meat, and 7) effects of broiler transport container sanitation and placement on Salmonella transmission. As part of this collaborative project, ARS researchers completed trials to determine the effects of broiler stunning method on meat biochemistry and functionality. As part of a separate ARS agreement, university collaborators conducted trials on alternative broiler slaughter and carcass chilling methods. A collaborative experiment was completed at a commercial broiler processing plant to investigate the effects of on-farm slaughter on product safety, quality, and processing efficiency. As part of this research, ARS researchers determined changes in meat biochemistry, functionality, and quality. In a separate pilot plant trial, ARS researchers investigated the effects of carcass position during delayed processing on broiler carcass and meat quality. More detailed descriptions of university collaborators⿿ research during FY2024 are included in separate annual performance reports for the agreements. Research progress on development of chemically modified non-selective agars for enhancement of contrast (Sub-objective 2A) and methodologies to eliminate semicarbazide production under poultry processing conditions as a means to eliminate production of semicarbazide in non-nitrofurazone treated poultry (Objective 3) were slowed due to a critical scientist vacancy. The recruitment process for filling this position was unsuccessful in FY2024 and is being restarted. Artificial Intelligence (AI)/Machine Learning (ML) AI methods used: In addition to various multimodal deep learning (DL) models combining convolutional neural networks (CNN) and long short-term memory (LSTM) networks, t-distributed stochastic neighbor embedding (t- SNE) was used for feature reduction and gradient-weighted class activation mapping was employed to investigate data transformations inside of the model. Finally, conformal prediction and deep k-nearest neighbor methods were adopted to measure and utilize the trustworthiness of bacterial detection models for unseen data points. Computing resources used: SCINet's HPC Clusters (Atlas) with graphical processing units were leveraged for computationally intensive DL involving large datasets and output. For less intensive computational tasks within the project, a local DL workstation was used, optimizing the utilization of available resources and ensuring efficient workflow management. Brief description of how the use of AI methods has benefited this project: While AI methods have led to the improved speed and accuracy for bacterial detection using automatic bacterial segmentation and detection based on multiple modalities, they also helped identify and reduce the generalization issue with the detection models. By employing explainable AI (XAI), feature reduction, and other AI techniques, we identified the root cause of the inconsistent accuracy in previous bacterial detection models when tested with external validation data, despite their high accuracy with internal validation data. This generalization issue was addressed by using uncertainty measurements of multiple modalities with a conformal prediction method for trustworthy ensemble. This approach significantly improved the detection accuracy with external validation data, increasing it from 40% to 89% while maintaining high detection accuracy with internal validation data. These findings suggest that even more generalizable ensemble models with higher accuracy can be developed from less-generalizable DL models of additional modalities using sound uncertainty measurements. ACCOMPLISHMENTS 01 Advances in generalizability of deep learning models for hypercubes from foodborne bacteria. Foodborne pathogens persist as a serious public safety concern in the United States impacting millions of people annually. While hyperspectral microscope imaging (HMI) combined with deep learning (DL) methods presents a potent strategy for the swift and accurate detection of foodborne bacteria, the widespread application of HMI-DL for food safety is somewhat constrained by generalization issues of the DL models for bacterial detection. ARS researchers at Athens, Georgia, developed an advanced artificial intelligence (AI) algorithm to identify and address a persistent problem with the generalizability of current AI models. This study delineated a method to account for the between-image variation that cause problematic spectral discrepancies across different datasets. This new AI approach improves data generalizability by eliminating the need for intricate per-image calibrations, a notable hurdle in the application of darkfield HMI technologies. Results showed that implementation of this correction strategy not only maintained bacterial detectability but also boosted the accuracy of the current Fusion-Net AI model from 38-70% to 95-99%. The enhanced AI model developed by ARS is a critical step towards the seamless integration of powerful HMI techniques into practical food safety investigations, marking a considerable advancement in foodborne pathogen detection.
Impacts
(N/A)
Publications
- Lu, Y., Jia, B., Yoon, S.C., Ni, X., Zhuang, H., Guo, B., Gold, S.E., Fountain, J.C., Glenn, A.E., Lawrence, K.C., Zhang, F., Wang, W., Lu, J., Wei, C., Jiang, H., Luo, J. 2024. Macro-micro exploration on dynamic interaction between aflatoxigenic Aspergillus flavus and maize kernels using Vis/NIR hyperspectral imaging and SEM technology. International Journal of Food Microbiology. 416. https://doi.org/10.1016/j.ijfoodmicro. 2024.110661.
- Hou, J., Park, B., Li, C., Wang, X. 2023. A multiscale computation study on bruise susceptibility of blueberries from mechanical impact. Postharvest Biology and Technology. https://doi.org/10.1016/j.postharvbio. 2023.112660.
- Huan, C., Shin, T., Park, B., Ro, K.S., Jeong, C., Jeon, H., Tan, P. 2024. Coupling hyperspectral imaging with machine learning algorithms for detecting microplastics in soils. Journal of Hazardous Materials. https:// doi.org/10.1016/j.jhazmat.2024.134346.
- Li, D., Park, B., Chen, Q., Ouyang, Q., Kang, R. 2024. Quantitative prediction and visualization of matcha color physicochemical indicators using hyperspectral microscope imaging technology. Food Control. https:// doi.org/10.1016/j.foodcont.2024.110531.
- Kang, R., Sun, S., Ouyang, Q., Huang, J., Park, B. 2024. 3D-GhostNet: A novel spatial-spectral algorithm to improve foodborne bacteria classification coupled with hyperspectral microscopic imaging technology. Sensors and Actuators B: Chemical. https://doi.org/10.1016/j.snb.2024. 135706.
- Campos, R., Yoon, S.C., Chung, S., Bhandarkar, S.M. 2023. Semisupervised deep learning for the detection of foreign materials on poultry meat with near-infrared hyperspectral imaging. Sensors. 23(16):7014. https://doi.org/ 10.3390/s23167014.
- Reina, M.A., Mcconnell, A., Figueroa, J.C., Riggs, M.R., Buhr, R.J., Price, S.B., Macklin, K.S., Bourassa, D.B. 2023. Quantification of Salmonella Infantis transfer from transport drawer flooring to broiler chickens during holding. Poultry Science. 103(2). Article 10377. https://doi.org/10. 1016/j.psj.2023.103277.
- Reina, M.A., Urrutia, A., Figueroa, J.C., Riggs, M.R., Macklin, K.S., Buhr, R.J., Price, S.B., Bourassa, D.B. 2023. Application of pressurized steam and forced hot air for cleaning broiler transport container flooring. Poultry Science. 103(2). Article 103276. https://doi.org/10.1016/j.psj. 2023.103276.
|
Progress 10/01/22 to 09/30/23
Outputs
PROGRESS REPORT Objectives (from AD-416): 1. Develop imaging technologies to detect and identify plastics during poultry processing with hyperspectral imaging and artificial intelligence. 1A. Develop hyperspectral imaging technology for detection and identification of plastic foreign objects during poultry processing. 1B. Develop AI technology for enhanced detection and smart robotic removal of foreign materials in hyperspectral imagery during poultry processing. 2. Detection and identification of foodborne bacteria and toxins in poultry products with high-throughput hyperspectral microscopy and surface plasmon resonance imaging. 2A. Rapid monitoring of indicator microorganisms in poultry processing. 2B. Develop advanced hyperspectral microscope imaging (HMI) methods and system for label-free detection and identification of pathogens at the cellular level with no enrichment. 2C. Develop high-sensitive and selective immunoassay method and system for foodborne bacteria and toxin detection with surface plasmon resonance imaging. 3. Eliminate the production of semicarbazide in non-nitrofurazone treated poultry by optimization of antimicrobial treatments and/or alternative antimicrobials during processing. 4. Develop safe and effective poultry processing strategies (scalding- picking-evisceration procedures) to reduce foodborne contaminants (pathogens/chemical) and enhance the sustainability of poultry processing. 4a. Develop sustainable poultry processing using artificial intelligence (AI) technology to improve poultry food safety. 4b. Develop Internet of Things (IoT) technology with various sensing platforms and data analytics for smart poultry processing and safety. Approach (from AD-416): Research on poultry safety will focus on: 1) developing and validating early, rapid, sensitive, and/or high-throughput optical sensing techniques for detecting physical and biological hazards in poultry products, and 2) eliminating semicarbazide in non-nitrofurazone treated poultry by optimization of antimicrobial treatments. In research on optical detection of physical hazards, spectroscopic and hyperspectral imaging (HSI) technologies will be developed for detection and identification of plastic foreign objects. A robot rejector and control software will be developed to eliminate foreign materials (FM) when detected by HSI. Artificial intelligence (AI) technology will be developed for enhanced detection and smart robotic removal of FM during poultry processing through the development and evaluation of customized deep learning algorithms based on hyperspectral imaging. A vision-guided smart robotic manipulator will be designed and built to remove FM by self- learning AI algorithms. To develop methods and techniques for detecting and identifying biological hazards, time-lapse image data on pure-culture indicator organisms and poultry carcass rinses from different processing locations will be collected to build a library, which will be used for on- line counting of microcolonies to build prototype systems. To detect foodborne pathogens, hyperspectral microscope imaging (HMI) methods will be developed with a spectral library of various pathogens using two HMI platforms of acousto-optical tunable filter (AOTF) and Fabry-Perot interferometer (FPI). In accordance with optimization of parameters on HMI and hypercubes, a transportable HMI system will be developed embedded with AI-based software for classification and identification. To identify foodborne bacteria and toxins, a highly-sensitive and selective immunoassay method and system will be developed using surface plasmon resonance imaging (SPRi). Microfluidic devices will be designed, simulated and fabricated for bacterial enrichment and separation. Both materials and parameters to develop a 3D printed biosensor for multiplex detection of pathogenic bacteria and toxins will be optimized and evaluated with food samples. Finally, a portable 3D printing platform for biosensor fabrication by integrating sample enrichment cartridge, biochip and SPRi detector will be developed. To develop techniques for eliminating the production of semicarbazide (SEM) in non-nitrofurazone treated poultry, a methodology for SEM analysis in chicken meat and a data library relating poultry processing conditions to SEM formation will be developed. Specifically, SEM in chicken leg quarters obtained from multiple processing facilities will be analyzed and methods to eliminate SEM production in poultry products under processing conditions will be developed. Research was conducted to develop hyperspectral imaging technology for detection and identification of plastic foreign objects during poultry processing (Sub-Objective 1A). Researchers made significant progress on an engineering project to develop a prototype imaging system for the detection and classification of plastic foreign materials on poultry meat by acquiring a near-infrared pushbroom hyperspectral camera producing up to 670 image frames/second and successfully constructing a race-track- like conveyor system with a closed loop. The closed loop conveyor system was specifically designed to enable continuous sample traveling (including chicken parts and foreign materials) for imaging purposes, utilizing the capabilities of the high-speed near-infrared hyperspectral camera to achieve real-time image acquisition and processing. Furthermore, progress was made on the design and in-house fabrication of a camera mount tailored for the prototype system. These system components will be integrated into the final design of the imaging system prototype. Progress was made on research to develop an artificial intelligence (AI) technology for enhanced detection and smart robotic removal of foreign materials in hyperspectral imagery during poultry processing (Sub- objective 1B). A real-time hyperspectral imaging (HSI) technology, incorporating high-performance deep learning (DL), is being developed to detect foreign materials, including plastics, on chicken breast meat. The challenge lies in achieving online real-time sensing capabilities critical for industrial deployment, given the computational demands of both HSI and DL. Researchers addressed this challenge by implementing hardware and software suitable for real-time HSI-based DL inferencing. To meet the processing time constraint of approximately 0.25 second/chicken breast fillet, the software system was designed to adopt parallel computing in both CPU and GPU, allowing for the parallelization of hyperspectral data acquisition and processing. Work done in FY2023 showed this approach enhanced the efficiency and speed of data handling. Researchers worked to design the system so that acquired and pre- processed hyperspectral data in the CPU are transferred to a GPU for foreign material detection using a previously developed DL model that employs CPU threads to mitigate transfer latency through asynchronous creation of multiple streams and overlapping multiple transfers with kernel execution. Within the GPU, tensor cores are then utilized to perform inference on the DL-based model trained to identify spectral responses associated with foreign materials. These advancements to address high computational demands will enable real-time detection of foreign materials, leveraging the power of HSI and DL technologies in the safety assessment of chicken meat products. Significant progress was made on a project on rapid monitoring of indicator microorganisms in poultry processing (Sub-objective 2A). Multiple time-lapse recordings were made of indicator microorganisms (non- pathogenic generic E. coli, Listeria, and Pseudomonas) growing on brain heart infusion (BHI) agar and standard methods agar (SMA). Three incubators were modified to capture high-resolution images (45.7MP) using a digital color camera at one-minute intervals over a typical 24 h incubation period. A total of 14 time-lapse recordings were completed, including repetitions of each bacterium on both agars. Change detection analyses were performed, resulting in binary images showing significant changes relative to the initial state image. Color time-lapse videos were created, and statistical means were computed and plotted against the timeline, providing a quantitative view of colony growth. Similar analyses were performed on Shiga toxin-producing E. coli (STEC) on chromogenic Rainbow agar, resulting in 15 additional data sets. The preliminary results showed that the average time for detecting microcolony growth and counting colonies was approximately 7, 14, and 15 h after incubation on BHI for generic E. coli, Listeria, and Pseudomonas, respectively. On SMA, average time was 9 h for generic E. coli and 16 h for Listeria. A low-angle darkfield illuminator was tested and found to offer higher contrast for future imaging purposes where bacterial color is not crucial. Significant progress was made to develop advanced AI methods for hyperspectral data analytics for label-free detection and identification of pathogens at the cellular level with no enrichment (Sub-objective 2B) using three approaches. First, research was conducted on bias removal from spectral data of pathogenic bacteria. The use of hyperspectral microscope imaging (HMI) in conjunction with deep learning (DL) has proven effective and efficient in identifying pathogenic bacterial cells. However, it was observed that models trained on HMI datasets collected in different years were spectrally biased, limiting their generalizability. To address this issue, ARS researchers worked to identify the source of bias and developed AI methods to remove bias while preserving bacterial detection capability. Following bias removal, the accuracy of AI classification models improved from 60% to 98%. This bias-removal method not only enabled model building for accurate and robust bacterial detection using a large amount of existing data but also allowed its application to new HMI datasets without the need for per-image calibration, which is particularly challenging for image acquisition using HMI systems. Second, incremental learning for efficient, accurate, and robust detection of foodborne bacteria was researched. Given the significant variation in spatial-spectral features of single-cell bacteria, it becomes crucial to update the detection model incrementally to adapt to practical data collection schedules and reduce computational memory requirements for big data during model training. Researchers at ARS also devised a new method for chunk-based incremental learning. This approach involves storing a minimal number of previous instances to retrain and update the model effectively with a new chunk of data. By employing this method and utilizing less than 40% of previous instances, the Fusion-Net model achieved a 99% test accuracy. When combined with the bias removal method, this incremental learning approach can further reduce the percentage of previous instances required for updating the model to detect foodborne bacteria. Third, model evaluation and redesign for bacterial detection using explainable AI were investigated. Through the utilization of explainable AI methods, ARS researchers discovered two issues with the existing training methods of the highly accurate Fusion- Net model for bacterial detection. Firstly, one of its subnetworks was not fully trained, and secondly, another subnetwork did not contribute significantly to overall model performance. To overcome these challenges, ARS researchers removed the non-contributing subnetwork from the Fusion- Net architecture followed by developing an algorithm that enabled simultaneous and complete training of all subnetworks while incorporating performance measurements for each subnetwork. The resulting Fusion-Net model exhibited similar high accuracy as before in bacterial detection but demonstrated enhanced robustness by utilizing all spatial-spectral features in HMI. This redesigned model was successfully employed for bacterial viability detection, achieving 100% accuracy, and species classification with a 98% test accuracy. During FY2023 significant progress was made to fabricate and test microfluidic devices for rapid detection of foodborne bacteria, such as Escherichia coli and Salmonella (Sub-objective 2C). Compared to conventional cultivation and molecular biology-based methods, microfluidics-based methods provide more freedom and flexibility for manufacturing biosensors in the lab and conducting rapid bacteria detection. ARS researchers worked with an external collaborator to conduct research on the fabrication of microfluidic devices for collecting foodborne bacteria. The microchannels were fabricated using two different technologies, 1) stereo-lithography (SLA) 3D printing and 2) computer numerical control (CNC) milling. To test the particle separation performance, fluorescent particles with different diameters were used to mimic bacteria and interfering particles in food matrices. At the same time, ARS researchers made progress in building a new fabrication platform based on state-of-the-art micro/nano 3D printing technology. Research progress on optimizing antimicrobial treatments and/or alternative antimicrobials during processing as a means to eliminate production of semicarbazide in non-nitrofurazone treated poultry (Objective 3) was slowed due to a critical scientist vacancy. The recruitment process for filling this position was initiated. Research to develop safe and effective poultry processing strategies to reduce foodborne contaminants and enhance the sustainability of poultry processing (Objective 4) was initiated. As part of a formal ARS agreement, university collaborators conducted research to determine the efficacy of utilizing controlled atmosphere stunning (CAS) in broiler slaughter as a means to enhance processing sustainability. Trials were conducted to determine the effects on food safety and processing efficiency. As part of this research effort, ARS researchers completed an initial trial to determine the interacting effects of the collaborator⿿s CAS system and different carcass deboning times on poultry meat quality and functionality. Artificial Intelligence (AI)/Machine Learning (ML) Artificial Intellegence (AI) methods used: 1. A novel semi-supervised hyperspectral deep learning (DL) model based on a generative adversarial network (GAN), with an improved 1D U-Net serving as its discriminator was developed. 2. To ensure transparency and understanding in bacteria detection, explainable AI methods were developed to validate and examine trained deep learning models with several feature reduction techniques including principal component analysis, partial least squares, t-distributed stochastic neighbor embedding, and uniform manifold approximation and projection for visualization of spectral data to identify data patterns for robust bacterial detection. Computing resources used: 1. Used external computing resources (a laptop computer with a GPU card) provided by an external collaborator at the University of Georgia, Athens, Georgia through a Research Support Agreement. 2. For computationally intensive tasks involving large datasets and output, we leveraged SCINet's HPC Clusters (Atlas). In addition to SCINet, we also utilized local computers for other computational tasks within the project, optimizing the utilization of available resources and ensuring efficient workflow management. Brief description of how the use of AI methods has benefited this project: 1. The developed semi-supervised AI method played a crucial role in expediting the research process for detecting foreign materials on poultry meat with enhanced accuracy, surpassing the performance of previously developed machine learning (ML) methods. Moreover, the utilization of semi-supervised learning allowed for leveraging a substantial amount of readily available unlabeled chicken meat data, which is typically more accessible and cost-effective to obtain compared to collecting labeled foreign material data. 2. The utilization of AI methods has brought numerous benefits to the preprocessing and multi-modal modeling of hyperspectral microscope imaging (HMI) data, resulting in improved accuracy for rapid bacteria detection. With AI-enabled auto-segmentation of bacterial single cells, the manual crafting of region of interest (ROI) images was eliminated, significantly accelerating the bacterial detection process. Our multi- modal Fusion-Net combined band images and spectra of single cells, enabling highly accurate identification of live bacterial cells. AI methods have also facilitated quality control, model optimization, and validation processes. Moreover, by utilizing feature extraction methods, we were able to identify the precise sources of bias in bacterial HMI. This led to the development of an effective method to remove bias from existing and/or new data, while preserving the capability for bacterial detection. Based on the new bias-removal method, implementing a new incremental learning method allows us to effectively remove bias from new data and identify valuable data points for efficient and effective learning for bacterial detection. Overall, AI methods have played a crucial role in extracting valuable information from the spatial-spectral dimensions of hyperspectral microscope images for foodborne bacterial detection. ACCOMPLISHMENTS 01 Developed Surface-enhanced Raman Spectroscopy (SERS) method with nanoparticle substrates for Salmonella detection in chicken rinse. Salmonella is a foodborne pathogenic bacteria commonly found on broiler chickens during processing that is responsible for causing gastrointestinal illnesses. ARS researchers in Athens, Georgia, developed a Salmonella detection method that reduces the necessary time for confirmation, by collecting Surface-Enhanced Raman Spectroscopy (SERS) spectra from bacteria colonies, applied to a substrate of biopolymer encapsulated silver nanoparticles. Chicken rinses containing Salmonella Typhimurium (ST) were analyzed by SERS and compared to traditional plating and polymerase chain reaction (PCR) methods. Analyses showed that SERS spectral features of ST and non-Salmonella colonies were significantly different. A support vector machine (SVM) classification algorithm was able to separate ST and non-Salmonella samples with an overall classification accuracy of 96%. Findings suggest that SERS is a highly accurate tool for pathogen detection that may be useful for regulatory purposes. 02 Automated segmentation of foodborne bacteria from chicken rinse with hyperspectral microscope imaging and deep learning methods. Foodborne illness is a significant threat to food safety and public health. A leading cause of foodborne illness is food contamination with pathogenic bacteria. It is crucial to identify pathogenic bacteria in contaminated food as early as possible. Hyperspectral microscope imaging (HMI) utilizes spectral-spatial features to identify pathogenic bacteria with a high level of accuracy. However, bacterial detection with dark-field HMI requires accurate segmentation of single-cell bacteria from hyperspectral images. ARS researchers in Athens, Georgia, developed a method to automatically segment single-cell pathogenic bacteria using deep learning and image processing for bacterial segmentation to identify a single-cell. To validate a method Escherichia coli, Listeria, Salmonella, and Staphylococcus were used to acquire hyperspectral imaging (hypercube) of bacterial cells under different growth conditions. Based on the hypercube, four different deep learning models were developed and evaluated for bacterial cell segmentation. AI-based automated segmentation methods performed with over 94% accuracy. This accurate and robust auto-segmentation technique streamlined the detection of pathogenic bacteria with HMI by reducing processing time from raw image acquisition to classification within 15 seconds. 03 Classification between live and dead foodborne bacteria with machine learning. Identification of live foodborne bacteria is essential for ensuring food safety and preventing foodborne illnesses. Accurate techniques for assessing bacteria viability are needed. ARS researchers in Athens, Georgia, developed deep learning methods for hyperspectral data analytics to accurately distinguish between live and dead foodborne bacteria based on their spectral and morphological features. Three deep learning models (Fusion-Net I, II, and III) were developed and evaluated for their ability to classify live and dead bacterial cells of six pathogenic strains, including Escherichia coli (EC), Listeria innocua (LI), Staphylococcus aureus (SA), Salmonella Enteritidis (SE), Salmonella Heidelberg (SH), and Salmonella Typhimurium (ST). Fusion-Net I achieved high accuracy in identifying live bacterial cells, with a classification accuracy of 100% for LI, SE, ST strains and over 92% for EC, SA, SH. Fusion-Net II and III models were even more robust, achieving 100% accuracy consistently in classifying dead cells in all six strains. Fusion-Net III also showed the ability to identify bacterial strains over 96% accuracy, making it a dual-task model with potential applications for early detection of live foodborne bacteria prior to outbreaks. These findings suggest that the use of hyperspectral microscope imaging and deep learning models could provide a new tool for identifying bacterial viability quickly and accurately, thereby improving the efficiency and reliability of food safety inspection. 04 Design of microfluidic devices for biosensor systems. Some of the major challenges of rapid bacteria detection in food matrices include: 1) limited sensitivity and specificity, and 2) high-throughput capacity to detect multiple bacteria at the same time. To solve these two problems, ARS researchers in Athens, Georgia, developed passive microfluidic devices which can separate, enrich, and detect bacteria from food matrices without applying incubating processes to the samples prior to detection. The passive microfluidic designs simplify the overall biosensor structure and make it easy to operate. Designs of microfluidic devices were achieved by fluidic dynamic simulations with various channel geometries, finding out the optimized microchannel sizes and geometries for the separation and enrichment of bacteria size particles from larger and smaller interfering particles in sample solutions. The passive microfluidic designs were useful to integrate multiple functions and high-throughput capacity into a simple and streamlined biosensor system to detect foodborne bacteria without conventional enrichment process. 05 Semicarbazide in chicken leg quarters obtained from multiple processing facilities. In recent years, the presence of semicarbazide in poultry has been confounding the U.S. poultry industry and causing export restrictions. Semicarbazide is a regulatory marker for the use of nitrofurazone, an antibiotic banned from use in animals intended for human consumption. ARS researchers in Athens, Georgia, conducted a survey of commercial processing plants to identify potential sources of semicarbazide increases in poultry during processing. Data indicated that semicarbazide increased during processing in some plants but not others. For those plants demonstrating an increase in semicarbazide during processing, chill tank conditions were identified as the primary source of processing induced semicarbazide formation. Further research is being conducted to assess chill tank parameters necessary to minimize the chemical production of semicarbazide during processing, primarily by pH regulation. These data are critical to the U.S. poultry industry to provide a basis to reopen export markets.
Impacts
(N/A)
Publications
- Park, B., Shin, T., Cho, J., Lim, J., Park, K. 2022. Improving blueberry firmness classification with spectral and textural features of microstructures using hyperspectral microscope imaging and deep learning. Postharvest Biology and Technology. https://doi.org/10.1016/j.postharvbio. 2022.112154.
- Eady, M.B., Setia, G., Park, B., Wang, B., Sundaram, J. 2023. Biopolymer encapsulated AgNO3 nanoparticle substrates with surface-enhanced Raman spectroscopy (SERS) for Salmonella detection from chicken rinse. International Journal of Food Microbiology. https://doi.org/10.1016/j. ijfoodmicro.2023.110158.
- Wang, Q., Childree, E., Box, J., Lopez-Vela, M., Sprague, D., Cherones, J., Higgins, B. 2023. Microalgae can promote nitrification in poultry- processing wastewater in the presence and absence of antimicrobial agents. ACS ES&T Engineering. https://doi.org/10.1021/acsestengg.2c00360.
- 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.
- Park, B., Shin, T., Wang, B., Mcdonogh, B., Fong, A. 2023. Classification between live and dead foodborne bacteria with hyperspectral microscope imagery and machine learning. Journal of Microbiological Methods. https:// doi.org/10.1016/j.mimet.2023.106739.
|
Progress 10/01/21 to 09/30/22
Outputs
PROGRESS REPORT Objectives (from AD-416): 1. Develop imaging technologies to detect and identify plastics during poultry processing with hyperspectral imaging and artificial intelligence. 1A. Develop hyperspectral imaging technology for detection and identification of plastic foreign objects during poultry processing. 1B. Develop AI technology for enhanced detection and smart robotic removal of foreign materials in hyperspectral imagery during poultry processing. 2. Detection and identification of foodborne bacteria and toxins in poultry products with high-throughput hyperspectral microscopy and surface plasmon resonance imaging. 2A. Rapid monitoring of indicator microorganisms in poultry processing. 2B. Develop advanced hyperspectral microscope imaging (HMI) methods and system for label-free detection and identification of pathogens at the cellular level with no enrichment. 2C. Develop high-sensitive and selective immunoassay method and system for foodborne bacteria and toxin detection with surface plasmon resonance imaging. 3. Eliminate the production of semicarbazide in non-nitrofurazone treated poultry by optimization of antimicrobial treatments and/or alternative antimicrobials during processing. 4. Develop safe and effective poultry processing strategies (scalding- picking-evisceration procedures) to reduce foodborne contaminants (pathogens/chemical) and enhance the sustainability of poultry processing. Approach (from AD-416): Research on poultry safety will focus on: 1) developing and validating early, rapid, sensitive, and/or high-throughput optical sensing techniques for detecting physical and biological hazards in poultry products, and 2) eliminating semicarbazide in non-nitrofurazone treated poultry by optimization of antimicrobial treatments. In research on optical detection of physical hazards, spectroscopic and hyperspectral imaging (HSI) technologies will be developed for detection and identification of plastic foreign objects. A robot rejector and control software will be developed to eliminate foreign materials (FM) when detected by HSI. Artificial intelligence (AI) technology will be developed for enhanced detection and smart robotic removal of FM during poultry processing through the development and evaluation of customized deep learning algorithms based on hyperspectral imaging. A vision-guided smart robotic manipulator will be designed and built to remove FM by self- learning AI algorithms. To develop methods and techniques for detecting and identifying biological hazards, time-lapse image data on pure-culture indicator organisms and poultry carcass rinses from different processing locations will be collected to build a library, which will be used for on- line counting of microcolonies to build prototype systems. To detect foodborne pathogens, hyperspectral microscope imaging (HMI) methods will be developed with a spectral library of various pathogens using two HMI platforms of acousto-optical tunable filter (AOTF) and Fabry-Perot interferometer (FPI). In accordance with optimization of parameters on HMI and hypercubes, a transportable HMI system will be developed embedded with AI-based software for classification and identification. To identify foodborne bacteria and toxins, a highly-sensitive and selective immunoassay method and system will be developed using surface plasmon resonance imaging (SPRi). Microfluidic devices will be designed, simulated and fabricated for bacterial enrichment and separation. Both materials and parameters to develop a 3D printed biosensor for multiplex detection of pathogenic bacteria and toxins will be optimized and evaluated with food samples. Finally, a portable 3D printing platform for biosensor fabrication by integrating sample enrichment cartridge, biochip and SPRi detector will be developed. To develop techniques for eliminating the production of semicarbazide (SEM) in non-nitrofurazone treated poultry, a methodology for SEM analysis in chicken meat and a data library relating poultry processing conditions to SEM formation will be developed. Specifically, SEM in chicken leg quarters obtained from multiple processing facilities will be analyzed and methods to eliminate SEM production in poultry products under processing conditions will be developed. During Fiscal Year (FY) 2022 significant progress was made on developing hyperspectral imaging technology for detection and identification of plastic foreign objects during poultry processing (Sub-Objective 1A). Hyperspectral images and spectra of common plastic materials such as polyethylene (PE), polyethylene terephthalate (PET), polyvinyl chloride (PVC), polystyrene (PS), polypropylene (PP), acrylonitrile butadiene styrene (ABS), and polyurethane (PUR) were collected and analyzed. An extended visible and near-infrared hyperspectral camera was used to collect hyperspectral images in the wavelength range 600 nm � 1,700 nm. A color camera and a spectrometer (400 nm � 2,500 nm) were also used. A classification technique was developed to identify the types of plastic pieces (foreign materials) found on breast fillets of broilers. During FY 2022 progress was made on developing artificial intelligence (AI) technology for enhanced detection and smart robotic removal of foreign materials in hyperspectral imagery during poultry processing (Sub- objective 1B). A deep learning AI model based on a generative adversarial network was developed for non-destructive, hyperspectral image-based foreign material detection, where no foreign materials were required at all during the training for the detection model while only chicken meat data were required for training. A dataset including about 900,000 spectral data were obtained from chicken breast fillets in the wavelength range of 1,000 � 2,500 nm. Tests were conducted for breast fillets contaminated with 30 different types of foreign materials commonly found in processing plants, in two sizes: nominal 2 mm x 2 mm (actual dimension of the longer side: 2.2 � 0.4 mm) and nominal 5 mm x 5mm (actual dimension of the longer side: 5.3 � 0.5 mm). The model achieved an F1 score of over 95% while the detection of foreign materials was 100%. During FY 2022 progress was made on the rapid monitoring of indicator microorganisms in poultry processing (Sub-objective 2A). Three dedicated incubators were modified for time-lapse imaging of colony growth in Petri dishes. Each incubator was modified to have a double-paned, sealed viewing port added to its top, through which a color camera can view and record bacterial growth in a single Petri dish. Each dish was back- illuminated by a specially modified light panel connected via fiber optic cable to an illuminator that resided outside of the incubator. Time-lapse images were captured at one-minute intervals at a resolution of 45.7 mega- pixels over a 48-hour period. The resulting high-resolution color images were assembled into a video at 8K resolution and analyzed using a change detection technique. Nine strains of Shiga toxin-producing E. coli have been imaged from serotypes O103, O121, O145, and O157. During FY 2022 significant progress was made to develop advanced hyperspectral microscope imaging (HMI) methods and systems for label-free detection and identification of pathogens at the cellular level with no enrichment (Sub-objective 2B) in three projects. First, research was conducted on multi-task AI detection of live pathogenic bacteria. Identifying live bacterial cells is critical to investigating potential foodborne outbreaks. ARS researchers developed a multi-task model to recognize live bacterial cells with hyperspectral microscope imaging and deep learning (HMI-DL) methods. The AI-based model classified bacterial cells as dead or alive, and identified their species including E. Coli, Listeria, Staphylococcus, and Salmonella with 100% and 96.9% test accuracies. In addition, the live-cell identification of the model was highly accurate because the model adjusted its parameters based on the self-predicted species. Second, explainable artificial intelligence (XAI) methods to identify dead bacterial cells were researched. Building a transparent model is crucial to understanding what the model learned from data. However, it is challenging to understand the input-output relationship with an AI-based model because such a model involves multiple layers of computations with its input. ARS researchers faced this lack of transparency with an AI model (Fusion-Net) identifying live bacterial cells with 100% test accuracy. To interpret the trained model, a visualization method called Gradient-weighted class activation mapping (Grad-CAM) was employed. This technique provided visual explanations of Fusion-Net's decision with given inputs and allowed the identification of spectral-spatial features influential for the AI model to identify live cells with hyperspectral microscope image data. Furthermore, the transparency of the model helped to condense structure and improve performance of the model. Third, deep learning methods for bacterial detection were investigated with USDA SCINet. Hyperspectral microscope imaging with deep learning (HMI-DL) has accurately detected pathogenic bacterial cells. But the technique requires heavy computation to train an AI-based model with its complex architecture. To reduce this computational burden, ARS researchers built a new deep learning environment in USDA SCINet. The USDA-ARS initiative provides a high- performance computing system with a hundred computers connected to a high- speed network. The developed deep learning environment utilized more than 5,000 processing cores and 30 GB of memory in the graphics processing units of the system, allowed accelerated training and continuous experimentation of the AI model to detect pathogenic bacterial cells with HMI. The SCINet environment improved the AI-based model (Fusion-Net) to classify four bacterial species (Salmonella, E. coli, Listeria, and Staphylococcus) with the combined big dataset of 940 GB collected over the four years (2019-2022) and classified the species with 98.4% test accuracy. During FY 2022 progress was made to develop a highly-sensitive and selective immunoassay method and system for foodborne bacteria and toxin detection with surface plasmon resonance imaging (Sub-objective 2C). Research on microfluidic sampling and biosensing systems for Escherichia coli and Salmonella was conducted. The development of portable biosensors for field-deployable detection has been increasingly important to control foodborne pathogens in the early stages of outbreaks. Since conventional cultivation and gene amplification methods require sophisticated instruments and highly skilled professionals, portable biosensing devices, which provide more flexibility for rapid detections even though their sensitivity and specificity are limited, are needed for high-throughput testing. Microfluidic methods have the advantage of miniaturizing instrumental size while integrating multiple functions and high- throughput capability into one streamlined system at low cost and minimal sample consumption to detect samples in different sizes and concentrations. ARS researchers investigated major active and passive microfluidic devices for bacteria sampling and biosensing focused on particle-based sorting/enrichment methods with or without external physical fields applied to the microfluidic devices for E. coli and Salmonella sampling. During FY 2022 progress was made on research to eliminate the production of semicarbazide in non-nitrofurazone treated poultry by optimization of antimicrobial treatments and/or alternative antimicrobials during processing (Objective 3). Research was conducted to develop a data library relating poultry processing conditions to formation of semicarbazide. The pH range necessary for the chemical production of semicarbazide from chicken treated with peracetic acid, as occurs in a poultry processing chill tank, has been plotted. ARS researchers have found that semicarbazide is produced in the pH range of 9 through 13. Semicarbazide production was not detected in the pH range 4 through 8. ACCOMPLISHMENTS 01 Development of a novel deep learning technique to detect foreign materials during poultry processing. Foreign materials found in poultry products are a food safety concern for consumers and poultry processors. ARS researchers in Athens, Georgia, developed a semi-supervised deep learning model based on a technique called generative adversarial network from hyperspectral images such that the need for collecting massive amounts of foreign material data for training is eliminated completely. Using this approach, results indicated that the detection of foreign materials of relatively small size (~ 2 mm x 2 mm) in poultry meat could be achieved with hyperspectral imaging at >95% accuracy. These findings demonstrate that implementing this deep learning AI model may make it possible to utilize hyperspectral imaging as an accurate, high-throughput system for foreign material detection during poultry processing. 02 Label-free immunoassay for multiplex detection of foodborne bacteria in chicken carcass rinse with surface plasmon resonance imaging. Frequent outbreaks of foodborne pathogens have stimulated the demand for biosensors capable of rapidly detecting multiple pathogens in contaminated food. ARS researchers in Athens, Georgia, developed sensing technology with surface plasmon resonance imaging (SPRi) for simultaneous label-free detection of multiple foodborne pathogens with a low limit of detection, mainly Salmonella spp. and Shiga-toxin producing Escherichia coli (STEC), in commercial chicken carcass rinse. The injected samples with different bacteria (Salmonella Enteritidis, STEC, and Listeria monocytogenes) have been identified by the same biochip. Moreover, the SPRi signals revealed complex interference effects among coexisting bacteria species in heterogeneous bacteria solutions. This SPRi-based immunoassay demonstrated great potential in high-throughput screening of multiple pathogenic bacteria coexisting in chicken carcass rinse. 03 Rapid identification of foodborne bacteria with hyperspectral microscopic imaging and artificial intelligent classification methods. Early detection of foodborne pathogens is crucial to promote public health. ARS researchers in Athens, Georgia, developed a technique called artificial intelligence (AI) assisted hyperspectral microscopic imaging (HMI) to differentiate five foodborne bacteria including Campylobacter, E. coli, Listeria, Salmonella, and Staphylococcus simultaneously. An artificial recurrent neural network called long- short term memory (LSTM) was employed and optimized to directly process the spectra acquired from different regions of interest of bacterial cells. Compared to conventional machine learning methods with classification accuracies between 66 - 85 %, the newly developed AI- based classifier achieved an accuracy of 92 %. Furthermore, the AI- assisted HMI system can predict spectra instantly, making it an efficient tool for foodborne bacteria identification. 04 Development of integrated and high-throughput bacteria sampling cartridge. Most biosensing instruments for bacteria detection and identification require purification and enrichment of samples. Microfluidic systems are capable of bacteria sampling with miniaturized devices. However, sensitivity and selectivity must be improved for pathogen control in real-world applications. ARS researchers in Athens, Georgia, developed a microfluidic cartridge for bacteria sampling that contains a compact structure of multiple stacked layers which is reusable and requires no additional reagent. This bacteria sampling cartridge can be embedded into benchtop instruments or coupled onto portable biosensors. From this research, an invention disclosure has been approved for a patent application and can be fabricated with 3D printing techniques that provides high spatial resolution and effective mechanical properties for microchannels inside the cartridge.
Impacts
(N/A)
Publications
- Yao, L., Beibei, J., Yoon, S.C., Zhuang, H., Ni, X., Guo, B., Gold, S.E., Fountain, J.C., Glenn, A.E., Lawrence, K.C., Zhang, H., Guo, X., Wang, W. 2022. Spatio-temporal patterns of Aspergillus flavus infection and aflatoxin B1 biosynthesis on maize kernels probed by SWIR hyperspectral imaging and synchrotron FTIR microspectroscopy. Food Chemistry. 382:132340. https://doi.org/10.1016/j.foodchem.2022.132340.
- Chung, S., Yoon, S.C. 2021. Detection of Foreign Materials on Broiler Breast Meat using Fusion of Visible Near-Infrared (VNIR) and Short-Wave Infrared (SWIR) Hyperspectral Imaging Modalities. Applied Sciences. https:/ /doi.org/10.3390/app112411987.
- Park, B., Shin, T., Cho, J., Lim, J., Park, K. 2021. Characterizing hyperspectral microscope imagery for classification of blueberry firmness with deep learning methods. Agronomy Journal. https://doi.org/10.3390/ agronomy12010085.
- Kang, R., Park, B., Ouyang, Q., Ren, N. 2021. Rapid identification of foodborne bacteria with hyperspectral microscope imaging and artificial intelligence classification algorithms. Food Control. https://doi.org/10. 1016/j.foodcont.2021.108379.
- Wang, B., Park, B. 2022. Microfluidic sampling and biosensing systems for foodborne Escherichia coli and Salmonella. Foodborne Pathogens and Disease. https://doi.org/10.1089/fpd.2021.0087.
- Wu, J., Ouyang, Q., Park, B., Kang, R., Wang, Z., Wang, L., Chen, Q. 2021. Physicochemical indicators coupled with multivariate analysis for comprehensive evaluation of matcha sensory quality. Food Chemistry. https:/ /doi.org/10.1016/j.foodchem.2021.131100.
- Zhang, H., Jia, B., Lu, Y., Yoon, S.C., Ni, X., Zhuang, H., Guo, X., Wang, W. 2022. Detection of aflatoxin B1 in single peanut kernels by combining hyperspectral and microscopic imaging technologies. Sensors. 22(13):4864. https://doi.org/10.3390/s22134864.
- Ouyang, Q., Wang, L., Park, B., Kang, R., Chen, Q. 2021. Simultaneous quantification of chemical constituents in matcha with visible near infrared hyperspectral imaging technology. Food Chemistry. https://doi.org/ 10.1016/j.foodchem.2021.129141.
- Park, B., Wang, B., Chen, J. 2021. Label-free immunoassay for multiplex detections of foodborne bacteria in chicken carcass rinse with surface plasmon resonance imaging. Foodborne Pathogens and Disease. http://doi.org/ 10.1089/fpd.2020.2850.
- Eady, M.B., Park, B. 2021. Unsupervised prediction model for Salmonella detection with hyperspectral microscopy: A Multi-year validation. Applied Sciences. https://doi.org/10.3390/app11030895.
- Mitchell, T.R., Glenn, A.E., Gold, S.E., Lawrence, K.C., Berrang, M.E., Gamble, G.R., Feldner, P.W., Hawkins, J.A., Miller, C.E., Olson, D.E., Chatterjee, D., Mcdonough, C.M., Pokoo-Aikins, A. 2022. Survey of meat collected from commercial broiler processing plants suggests low levels of semicarbazide can be created during immersion chilling. Journal of Food Protection. 85(5):798-802. https://doi.org/10.4315/JFP-22-012.
|
Progress 10/01/20 to 09/30/21
Outputs
Progress Report Objectives (from AD-416): 1. Develop imaging technologies to detect and identify plastics during poultry processing with hyperspectral imaging and artificial intelligence. 1A. Develop hyperspectral imaging technology for detection and identification of plastic foreign objects during poultry processing. 1B. Develop AI technology for enhanced detection and smart robotic removal of foreign materials in hyperspectral imagery during poultry processing. 2. Detection and identification of foodborne bacteria and toxins in poultry products with high-throughput hyperspectral microscopy and surface plasmon resonance imaging. 2A. Rapid monitoring of indicator microorganisms in poultry processing. 2B. Develop advanced hyperspectral microscope imaging (HMI) methods and system for label-free detection and identification of pathogens at the cellular level with no enrichment. 2C. Develop high-sensitive and selective immunoassay method and system for foodborne bacteria and toxin detection with surface plasmon resonance imaging. 3. Eliminate the production of semicarbazide in non-nitrofurazone treated poultry by optimization of antimicrobial treatments and/or alternative antimicrobials during processing. Approach (from AD-416): Research on poultry safety will focus on: 1) developing and validating early, rapid, sensitive, and/or high-throughput optical sensing techniques for detecting physical and biological hazards in poultry products, and 2) eliminating semicarbazide in non-nitrofurazone treated poultry by optimization of antimicrobial treatments. In research on optical detection of physical hazards, spectroscopic and hyperspectral imaging (HSI) technologies will be developed for detection and identification of plastic foreign objects. A robot rejector and control software will be developed to eliminate foreign materials (FM) when detected by HSI. Artificial intelligence (AI) technology will be developed for enhanced detection and smart robotic removal of FM during poultry processing through the development and evaluation of customized deep learning algorithms based on hyperspectral imaging. A vision-guided smart robotic manipulator will be designed and built to remove FM by self- learning AI algorithms. To develop methods and techniques for detecting and identifying biological hazards, time-lapse image data on pure-culture indicator organisms and poultry carcass rinses from different processing locations will be collected to build a library, which will be used for on- line counting of microcolonies to build prototype systems. To detect foodborne pathogens, hyperspectral microscope imaging (HMI) methods will be developed with a spectral library of various pathogens using two HMI platforms of acousto-optical tunable filter (AOTF) and Fabry-Perot interferometer (FPI). In accordance with optimization of parameters on HMI and hypercubes, a transportable HMI system will be developed embedded with AI-based software for classification and identification. To identify foodborne bacteria and toxins, a highly-sensitive and selective immunoassay method and system will be developed using surface plasmon resonance imaging (SPRi). Microfluidic devices will be designed, simulated and fabricated for bacterial enrichment and separation. Both materials and parameters to develop a 3D printed biosensor for multiplex detection of pathogenic bacteria and toxins will be optimized and evaluated with food samples. Finally, a portable 3D printing platform for biosensor fabrication by integrating sample enrichment cartridge, biochip and SPRi detector will be developed. To develop techniques for eliminating the production of semicarbazide (SEM) in non-nitrofurazone treated poultry, a methodology for SEM analysis in chicken meat and a data library relating poultry processing conditions to SEM formation will be developed. Specifically, SEM in chicken leg quarters obtained from multiple processing facilities will be analyzed and methods to eliminate SEM production in poultry products under processing conditions will be developed. This project replaces project 6040-42000-044-000D, "Develop Rapid Optical Detection Methods for Food Hazards," which ended March 2021. Substantial progress has been made on this project. Hyperspectral imaging technology for detection and identification of plastic foreign objects: In support of Objective 1, ARS researchers in Athens, Georgia, collected and analyzed spectra and hyperspectral images of 30 different types of foreign materials including plastic, rubber gloves, metal, fabric, and wood with wavelength range from 400 nm to 2,500 nm twice. A sensor fusion technique was developed to use spatial and spectral information of two hyperspectral imaging systems operating in two different spectral ranges of visible and near-infrared (400 nm-1,000 nm and 1,000 nm-2,500 nm). Spectral-spatial deep learning techniques for detection of Shiga toxin producing Escherichia coli (STEC) colonies on agar plates: In support of Objectives 1 and 2, ARS researchers in Athens, Georgia, developed a deep learning technique using spatial-spectral features in hyperspectral images to detect 15 different STEC serovars on solid agar media. Both modified MacConkey and modified Rainbow agar types were compared for the study. The overall classification accuracies were 91% and 94% for the MacConkey and Rainbow agar, respectively. Early detection of microcolonies of indicator microorganisms in poultry processing: In support of Objective 2, ARS researchers in Athens, Georgia, constructed an incubation system allowing time-lapse imaging of agar plates in situ to study the growth of three non-pathogenic bacteria (Pseudomonas putida, Listeria innocua, and Escherichia coli K12) on non- selective agar plates. The study results suggested that colonies were detectable with a high-resolution digital camera as early as 8 hours. Development of automatic bacteria cell segmentation methods: Single-cell bacteria segmentation has been a bottleneck for rapid bacterial detection with hyperspectral microscope imaging (HMI) because low-quality results from existing automatic segmentation methods prevented them from practical use and high-quality output from manual segmentation required a good amount of expert's time and effort (e.g., 30-60 minutes per hyperspectral data or hypercube). In support of Objective 2, ARS researchers in Athens, Georgia, developed an automatic segmentation method that consisted of two steps, bacterial segmentation with deep learning (DL) and single-cell selection with ellipse fitting evaluation. A DL-based segmentation method performed with 94.1% accuracy and less than 15 seconds for the result, which was better than previous manual segmentation methods with 88% and up to 60 minutes. Design and simulation of microfluidic channels for bacteria separation and enrichment: In support of Objective 2, ARS researchers in Athens, Georgia, developed simulation models to enrich Salmonella with six different sizes of particles (500 nm, 1 �m, 2 �m, 3 �m, 10 �m, and 20 �m) to represent various species in complex food matrix for three types of microfluidic channels. The channel width, height, aspect ratio of cross- section, and injection flow rate were optimized. The simulation results on Salmonella concentration and the microfluidic channel design parameters for optimum separation/enrichment efficiency were obtained. Evaluation of Poultry Processing Conditions for the Formation of Semicarbazide (SEM) on Chicken Products: In support of Objective 3, collaborative research was conducted to evaluate poultry processing conditions that may result in the formation of semicarbazide (SEM) on chicken products. A comprehensive survey of broiler processing establishments was continued to get a better understanding of the formation of SEM on chicken products to determine which factors may or may not contribute to this formation. Frozen leg quarters were received from 24 processing plants at several locations and shifts throughout the plant. Environmental factors such as antimicrobial treatment pH, temperature and time were recorded and submitted with each sample. Over the course of the project a total of 576 samples were received, separated, carefully logged, ground, and frozen in 2 g samples for analysis. Each sample was repeated in triplicate. A multi-step solvent extraction method was performed on each sample for future HPLC (High Performance Liquid Chromatography) analysis by collaborators. Record of Any Impact of Maximized Teleworking Requirement: Significant progress has been made on most of our research plan. However, maximized teleworking requirement hindered and delayed research progress that required data collection from experiments in laboratory. Especially, with regards to research on design and fabrication of microfluidic device for sampling foodborne pathogens in chicken rinse, no experiment has been conducted since maximized teleworking requirement. Due to COVID-19 and future uncertainties and limited access to laboratory and in-plant poultry production lines for on-site experiments, we are not certain our research goals can be accomplished as planned. Thus, we need to realign and reprioritized our research plan to accomplish the goal as much as we can under COVID-19 maximized teleworking circumstances. ACCOMPLISHMENTS 01 Artificial Intelligence (AI) classification of pathogens from chicken rinse. Food contamination with pathogenic bacteria is a leading cause of foodborne illness and requires early and rapid detection of pathogens in food matrices. Current detection and classification methods have limitations with regards to their implementation for field- deployable detection due to the high volume of samples that are needed for regulatory purposes. ARS researchers in Athens, Georgia, developed a method to detect foodborne pathogens in chicken rinse using hyperspectral microscope imaging and deep learning (HMI-DL) techniques. The developed model called Fusion-Net is an artificial intelligence algorithm that can identify single-cell bacteria in hyperspectral images. The model was trained with four species (Salmonella, E. coli, Listeria, and Staphylococcus) spiked in chicken rinse. The AI-based Fusion-Net model was able to classify foodborne pathogens with 98.7% test accuracy. The result proved that HMI-DL technique has the potential for rapid, high-throughput detection of foodborne bacteria at the single cell level. A commercial partner has adopted the ARS model and included it in their HMI interface operating software for bacteria classification.
Impacts
(N/A)
Publications
|
|