Deep Learning Enables Automatic Detection of Bacterial Swarming Based on A Single Image

Introduction: : Efficient detection of bacterial swarming motility and its differentiation from planktonic swimming motility is essential in biological research and the clinical diagnosis of diseases such as urinary tract infections (UTI) and inflammatory bowel disease (IBD). Both swarming and swimming are primary bacterial motility mechanisms driven by flagella; however, swimming involves more individualistic, free-form movement in liquid environments, whereas swarming is characterized by collective movement on semi-solid surfaces facilitated by flagella and surfactants. Traditional methods to distinguish swarming from swimming rely on qualitative, culture-based techniques that observe colony morphology. These approaches are time-consuming, lack specificity, and depend heavily on subjective assessments. Although recent advancements involving video-based analysis within small confinements at colony edges have significantly improved detection specificity and efficiency, they still require expert manual visual verification, making them labor-intensive and susceptible to human error. To overcome these limitations, this study introduces a deep learning-based swarming detection network that utilizes a single image with a long integration time as its input. This method enables the neural network to learn the encoded spatiotemporal features associated with bacterial motion on a single frame through motion blur, providing a rapid, automated, objective, and quantitative assessment of bacterial swarming.

Materials and

Methods: : To ensure clear differentiation and avoid label confusion, bacterial swarming (positive) and swimming (negative) samples were prepared using distinct experimental protocols according to established standards. After preparation, the samples were placed under a phase-contrast microscope to capture videos. For each observation well, image sequences were cropped into a 500×500 pixel region (~86×86 μm²) across all time points. Subsequently, every set of 10 consecutive frames (~0.3 seconds) was averaged, encoding the dynamic bacterial motion into a single integrated image. These integrated images, serving as network inputs, were then fed into a deep learning-based swarming classifier to determine bacterial swarming probability. The classifier employs a DenseNet architecture integrated with an attention module. It was trained and validated on a dataset comprising 1,301 positive images from 52 SM3 swarming wells and 2,703 negative images from 38 SM3 swimming wells. During blind testing, the classifier automatically computes the probability of swarming for unseen wells. Wells exhibiting a swarming probability greater than 0.72 were classified as positive—a threshold specifically determined through validation data to optimize sensitivity while maintaining 100% specificity.

Results, Conclusions, and Discussions:: Although trained exclusively on SM3 samples, our deep learning-based swarming classifier was evaluated on both unseen SM3 data for internal generalization and on two additional bacterial strains—DB10 (Serratia marcescens, laboratory strain) and H6 (Citrobacter koseri)—demonstrating external generalization. Specifically, on 39 swarming and 44 swimming well images from SM3, the classifier achieved a sensitivity of 97.44% and a specificity of 100%. When tested on DB10, it attained a sensitivity of 97.92% and a specificity of 96.77% across 48 swarming and 31 swimming images. For the H6 strain, the model achieved 100% sensitivity and 97.22% specificity on 27 swarming and 36 swimming images. These results demonstrate that the model’s learned features generalize robustly to new bacterial strains without the need for transfer learning.

In conclusion, the high accuracy and robustness of our deep learning model underscore its significant potential as a rapid, objective, and practical tool for on-site screening of bacterial swarming motility. This model offers promise for improving the early detection and monitoring of diseases associated with swarming motility, such as IBD and UTI.

Acknowledgements and/or References (Optional)::