Deep learning–enabled detection of bacterial swarming from a single image

We report an attention‐based deep convolutional neural network that autonomously predicts the probability of bacterial swarming from a single microscopic image with a long integration time. By capturing the spatiotemporal signatures of collective motion encoded in a stationary image, our classifier eliminates the need for extended video acquisition and manual feature extraction—delivering a rapid, automated, and quantitative assessment of bacterial motility. Trained on Enterobacter sp. SM3, the model achieved a sensitivity of 97.4% and a specificity of 100% in blinded SM3 evaluations, accurately distinguishing swarming (positive) from swimming (negative) phenotypes in new, unseen SM3 samples. Without retraining, the model generalized robustly to unseen bacterial species such as Serratia marcescens DB10 (97.9% sensitivity, 96.8% specificity) and Citrobacter koseri H6 (100% sensitivity, 97.2% specificity). Processing each image takes < 0.4s on standard GPU hardware, and this platform enables high-throughput microbial phenotyping, holding significant promise for on-site, point-of-care diagnostics of swarming-associated infections.