Classification of E. coli colony with generative adversarial networks, discrete wavelet transforms and VGG19

Authors: YADAV, PAPPU - University Of Florida; BURKS, THOMAS - University Of Florida; DUDHE, KUNAL - University Of Florida; FREDERICK, QUENTIN - University Of Florida; Qin, Jianwei - Tony Qin; Kim, Moon; RITENOUR, MARK - University Of Florida

Submitted to: Veterinary Radiology and Ultrasound
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 7/18/2023
Publication Date: 7/28/2023
Citation: Yadav, P., Burks, T.F., Dudhe, K., Frederick, Q., Qin, J., Kim, M.S., Ritenour, M.A. 2023. Classification of E. coli colony with generative adversarial networks, discrete wavelet transforms and VGG19. Veterinary Radiology and Ultrasound. 6(3):146-160.

Interpretive Summary: Fruits and vegetables (e.g., citrus) can host bacterial pathogens (e.g., E. coli) that can cause severe health issues for the consumers. Detection of bacterial colonies on fruits and vegetables is important to reduce food safety risks and foodborne diseases. This study developed a method for inspecting E. coli colonies using a contamination, sanitization inspection and disinfection (CSI-D) handheld fluorescence imaging device. Fluorescence images were collected from different concentrations of E. coli populations inoculated on black rubber slides. State-of-the-art deep learning algorithms (i.e., convolutional neural networks and generative adversarial networks) were used for image classifications. The best accuracy was achieved at 97% to classify four concentration levels of the E. coli colonies. The combination of the CSI-D handheld imaging and deep learning techniques would benefit the food industry and the regulatory agencies (e.g., USDA FSIS and FDA) in ensuring and enforcing the food safety standards for the products related to fruits and vegetables.

Technical Abstract: The transmission of Escherichia coli (E. coli) bacteria to humans through infected fruits, such as citrus, can lead to severe health issues, including bloody diarrhea and kidney disease (Hemolytic Uremic Syndrome). Therefore, the implementation of a suitable sensor and detection approach for inspecting the presence of E. coli colonies on fruits and vegetables would greatly enhance food safety measures. This journal article presents an evaluation of SafetySpect's Contamination, Sanitization Inspection, and Disinfection (CSI-D+) system, comprising an UV camera, an RGB camera, and illumination at two fluorescence excitation wavelengths: ultraviolet C (UVC) at 275 nm and violet at 405 nm. To conduct the study, different concentrations of bacterial populations were inoculated on black rubber slides, chosen to provide a fluorescence-free background for benchmark tests on E. coli-containing droplets. A VGG19 deep learning network was used for classifying fluorescence images with E. coli droplets at four concentration levels. Discrete wavelet transforms (DWT) were used to denoise the images and then generative adversarial networks (StyleGAN2-ADA) were used to enhance dataset size to mitigate the issue of overfitting. It was found that VGG19 with SoftMax achieved an overall accuracy of 84% without synthetic datasets and 94% with augmented datasets generated by StyleGAN2-ADA. Furthermore, employing RBF SVM increased the accuracy by 2% points to 96%, while Linear SVM enhanced it by 3% points to 97%. These findings provide valuable insights for the detection of E. coli bacterial populations on citrus peels, facilitating necessary actions for decontamination.