Deep Learning and Physics-Inspired Modeling for Cell Segmentation and Tracking with Application to Bacterial Biofilms

Synthetic cell image generation, as well as cell segmentation and tracking from microscopy, are essential in biology and biomedical research for advancing scientific understanding of a cell population. Synthetic image generation allows the testing of algorithms and the training of data-driven methods. The segmentation task allows for identifying, isolating, and analyzing individual cells from images, while tracking enables the analysis of cell behavior over multiple frames of an image sequence. Developing cell segmentation and tracking algorithms is an active research domain that propels drug discovery, disease diagnosis and treatment, tissue engineering, and basic research in cell biology.

In this dissertation, we introduce novel synthetic image generation, cell segmentation, and tracking algorithms to study a particularly challenging cell population called bacterial biofilms from lattice light-sheet microscopy 3D images and videos. Biofilms are complex biological systems that have critical functions in diverse fields, including the production of bioelectricity and the development of infectious diseases. High cell density and intra-cellular intensity inhomogeneity in the microscopy images of biofilms pose significant challenges to the existing algorithms in identifying individual bacterial cells and tracking their movements over time. The dissertation achieves three main objectives. (1)
We present a simulation framework designed to produce synthetic biofilm images and videos featuring cells with realistic curvilinear morphology. This framework is demonstrated to be useful for training supervised deep learning models for cell segmentation and tracking purposes. (2) We propose a novel deep learning-based cell segmentation approach that involves enhancing the cell interior and border information leveraging Euclidean distance transforms and then detecting cell seeds for a classical watershed segmentation through voxel-wise classification. (3) We introduce an innovative cell tracking framework that incorporates a deep temporal sequence classification network to predict the probability of potential associations between consecutive frames, followed by a one-to-one matching optimization to establish accurate matches. The tracking approach also considers the detection of cell division events via an Eigen decomposition-based strategy.