Date of Award
Doctor of Philosophy (PhD)
School of Computing
Marc R. Birtwistle
Brian C. Dean
Tracking cells over time is a fundamental task in live-cell imaging, and often requires costly manual analysis if images are not acquired with high enough frame rate. Acquiring high frame rate images, however, can limit the number of conditions explored and cells analyzed, and contribute to photobleaching, which makes fluorophores dimmer and phototoxicity, which affects cell health and renders the resulting data unusable.
Assuming a relatively high frame rate in image acquisition, state-of-the-art cell tracking approaches rely on either spatial proximity or morphological similarity to link cells in consecutive frames. The problem is that, at low frame rate, both approaches fall short since the position and appearance of cells can change significantly.
The goal of this thesis is to improve the robustness of cell tracking at low acquisition rate. To this end, we started by focusing on the first computational problem in cell tracking, which is cell identification. Convolutional Neural Networks (ConvNets) provide a way to accurately detect cells, but the manual annotation needed for training is costly. Thus, our first research question is focused on 1) how to train deep ConvNets for cell identification without manually-annotated cell labels? We proposed an image processing pipeline which uses fluorescent images to generate cell labels for training ConvNets. The experiment results showed that the proposed model can achieve competitive performance (recall 0.89 and precision 0.92) for identifying cells in a completely automatic manner. Then, we focused on the actual cell tracking problem, i.e., how to follow cells in consecutive frames. Inspired by the biologically proven theory that a cell's morphology suggests its moving direction, we studied 2) if we can design a set of features to represent the cell shape and estimate the cell velocity by regression to predict the cell's future position. We used hand-crafted geometric features for modeling the shape of cells and the experiments demonstrated that the cell velocity can be estimated using these features. Given that geometric features extracted from image patches can describe the motion of a cell, we focused on the third research question which is 3) how to integrate cell velocity estimations to improve the cell tracking accuracy at low frame rates? Our proposed approach contains two innovative components. First, we proposed a new deep-learning-based approach to automatically derive cell velocity information from image patches without the need for manually-defined geometric features. Second, we designed a new Bayesian framework which leverages cell position information and cell velocity estimations to track cells. We compared our cell linking method to both state-of-the-art tracking approaches and tracking algorithms implemented in well-established toolboxes for cell analysis. Our approach outperformed existing methods while allowing a 4x reduction in the frame rate.
Zhang, Xiang, "Cell Tracking at Low Frame Rate using Deep Learning and Bayesian Integration" (2022). All Dissertations. 3051.