Abstract
Early detection of foodborne pathogens is critical for improving food safety and preventing contamination-related outbreaks. Conventional methods for detecting Salmonella spp. are reliable but require extended incubation times and labor-intensive procedures, limiting their use for rapid screening. In this study, we developed a rapid and automated detection framework that integrates phase-contrast microscopy with a YOLOv8x-based deep learning model to identify Salmonella enterica serovar Typhimurium at the microcolony stage. A dataset of 5,000 microscopy images, including pure and mixed cultures with Escherichia coli K-12 under both clean and food-matrix (onion) conditions, was collected across multiple time points (30 min to 4 h). The model achieved a precision of 0.891, recall of 0.867, and mAP@0.50 of 0.898. Reliable detection was achieved as early as 2 hours of incubation, significantly reducing the time required compared to conventional culture-based methods. The model also demonstrated robust performance in mixed cultures, effectively distinguishing between morphologically similar bacterial species. In addition, instance-level analysis enabled quantitative characterization of microcolony growth dynamics over time. This approach provides a rapid, accurate, and cost-effective strategy for early-stage pathogen detection and has strong potential for real-time food safety monitoring and microbial diagnostics.