Motion-refined machine learning enables characterization of bacterial swarming dynamics.

Bacterial swarming on a semi-solid agar surface is a prevalent form of collective motility. Our study focuses on probing the swarm front of a novel species of gut bacteria, Enterobacter sp. SM3, which manifests strong swarming behavior. By depositing a fluid drop on the edge of the swarm, a monolayer of swarming bacteria is observed in minutes. We image these swarming bacteria and segment/identify them with the aid of CellPose, a machine-learning-based software algorithm. To address the challenge of dense, highly motile populations where automated segmentation alone is insufficient, we implement a post-processing approach that leverages particle image velocimetry (PIV) in conjunction with intersection-over-union (IoU) mapping to propagate segmentation masks across consecutive frames. This approach corrects spurious split and merge events introduced by automated segmentation, integrates a scientist-in-the-loop framework for targeted refinement of cell identities, and improves downstream trajectory reconstruction using TrackMate within Fiji. Together, the integration of machine-learning-based segmentation with PIV-IoU post-processing enables robust tracking of individual bacteria in crowded environments, allowing comprehensive analysis of their individual trajectories and collective dynamics.