A Deep Learning Framework for Tracking Motile Bacteria Leveraging Semi-Synthetic Image Augmentation and Quantitative Analysis of Run-and-Tumble Dynamics

Abstract

Swimming bacteria are widely studied as a representative model organism in active matter, providing a key experimental system for investigating single-cell self-propulsion and stochastic motility dynamics. In this thesis, I extend this line of research through two complementary studies. First, to quantitatively analyze swimming dynamics of moderately dense populations of rod-shaped bacteria, I develop a tracking pipeline that distinguishes partially overlapped bacterial cells using embedding-based instance segmentation trained solely on semi-synthetically augmented images, eliminating the need for manual labeling. By generating semi-synthetic training images that combine real cell image patches with diverse background conditions, I train an embedding-based instance segmentation model capable of reliably resolving individual cells even under partial overlaps and in complex imaging environments. The complete pipeline, released as an open-source software package named SynEmbTrack, is publicly available on GitHub. Second, using the reconstructed single-cell trajectories, I extract key motility features—including transition rates of run to tumble and tumble to run, run speed, and noises for speed and angular displacements— and analyze their temperature dependence across seven conditions (20◦C–50◦C). To infer the underlying motility parameters, I employ a sequential optimization strategy that evaluates a carefully defined loss function comparing experimental distributions and autocorrelations of the speed and orientation with stochastic run-and-tumble simulations. This approach yields consistent and robust estimates of the run-and-tumble parameters across temperatures, revealing clear temperature dependent trends in motility behavior. Together, these two studies provide an integrated framework that features image-based single-cell tracking and stochastic model–based parameter inference. The results establish a generalizable analytical tool for studying bacterial motility and offer insights into the microscopic mechanisms underlying temperature-dependent swimming dynamics. This framework also provides a basis for extending the analysis to more complex environmental conditions, including spatio-temporal gradients that induce taxis, as well as interactions with the surrounding medium or boundary geometries.