Microcolony-based instance segmentation framework for rapid quantification of antimicrobial susceptibility in Staphylococcus aureus

초록

Rapid evaluation of antimicrobial activity is essential for controlling microbial contamination and ensuring food safety. Recent advances in artificial intelligence have enabled the automatic detection of bacterial colonies from microscopic images, creating new opportunities for rapid microbial analysis. In this study, a YOLO-based instance segmentation model was developed to detect and quantify early-stage microcolonies of Staphylococcus aureus (S. aureus) using microscopic imaging. The trained YOLO model demonstrated reliable detection performance, with a precision of 0.93, a recall of 0.86, an mAP@0.5 of 0.92, and an mAP@0.5:0.95 of 0.70. It enables quantification of pixel-based bacterial microcolony detection. A calibration model trained on the total pixel area of YOLO-detected colonies showed a strong correlation with conventional plate-counting results (R2 = 0.9987). In addition, antimicrobial susceptibility testing was conducted using five inorganic antimicrobial compounds (AlCl3, ZnCl2, FeCl3, SiCl4, and TiCl4). The minimum inhibitory concentration and minimum lethal concentration values were determined to be 3150 and 3500 mg/kg for AlCl3, 2430 and 2700 mg/kg for ZnCl2, 2430 and 2700 mg/kg for FeCl3, 1620 and 1800 mg/kg for SiCl4, and 4500 and 5000 mg/kg for TiCl4, respectively. These results demonstrate that YOLO-assisted microcolony detection can reliably quantify S. aureus growth and evaluate antimicrobial activity at the microcolony stage under in vitro conditions. Although further studies using complex food matrices are needed before direct application in food systems, this approach can be a useful screening platform for rapid antimicrobial susceptibility testing related to food safety.

키워드