Deep learning enabled rapid detection of live bacteria in the presence of food debris
Hyeon Woo Park, Zhuo Li, Luyao Ma, Nitin Nitin
Abstract
The contamination of food with pathogenic bacteria is a major public health concern, requiring rapid and accurate detection methods. Conventional approaches, such as culture-based or molecular assays, are time-consuming, labor-intensive, and often demand specialized expertise. Here, we developed a deep learning-based strategy for rapid detection and classification of live bacteria using simple white-light microscopic images of microcolonies, even in the presence of morphologically similar food debris. The model, based on ResNet50 with a Region Proposal Network, was trained on Escherichia coli, Listeria monocytogenes, Bacillus subtilis, and debris from chicken, spinach, and cheese. The model trained on bacteria misclassified debris as bacteria (24.2% false positives), whereas the model trained on both bacteria and food debris achieved 0% false positives with 100% precision and 94.4% recall. Validation with GFP-producing B. subtilis in food matrices further confirmed robust performance (mPrecision 94.6%, mRecall 92.5%). This cost-effective method enables reliable bacterial detection in complex foods within 3 h.