Segmentation and Machine Learning for the Analysis of Bacterial Biofilm Images

As the first forms of life on earth billions of years ago, bacteria are essential participants in nature and exist almost everywhere. However, the living world of these cells remains mysterious due to limitations of conventional microscopy.

The recently introduced lattice light sheet microscope is capable of breaking the diffraction limit and performing long term live cell imaging with low photo-toxicity. This new modality necessitates advanced image analysis algorithms for understanding or ultimately controlling the activities of individual bacteria in a crowded, three-dimensional environment such as a biofilm. Bacterial behavior in biofilms is closely related to important problems in energy, disease, and the basic biology. Although there exist bacterial image analysis algorithms, they fail to delineate cells in dense biofilms, especially in 3D imaging scenarios in which the cells are growing and subdividing in a complex manner. This research develops an automated and effective analysis toolkit that overcomes the challenges of live cell imaging, such as the lack of apparent structure, limited contrast between conterminous cells, and high density of cells.

In the initial thrust of this research, a level set segmentation workflow, named Bact-3D, was explored with local constraints to stop the merging of level sets in different cell regions. The algorithm yielded promising results for multi-layered biofilm 3D data, but the performance degrades when applied to more complex biofilm images, where the gaps between neighboring cells are extremely hard to distinguish. Therefore, linear graph cuts (LCuts) is proposed as an automated cell segmentation algorithm to find individual bacteria by extracting and detecting the embedded linearity features in the biofilm. LCuts and its later improvements are generally extendable as a linear data clustering method; the method does not require prior knowledge of the number of cells as do other clustering methods. In the second major thrust of this dissertation, we investigate the incorporation of LCuts with deep neural networks to maximize the cell detection accuracy, and propose a generalized and unified algorithm, m-LCuts, for post-processing under- and over-segmented results.

Current training data are limited due to the limited number of real datasets and lack of corresponding annotated ground truth. The current gold-standard, manual annotation, is error-prone and time consuming. Therefore, as efforts to expand the annotated 3D biofilm datasets, both a model-based image simulation pipeline using optical and biological knowledge and an image generation workflow via 3D cyclic generative adversarial networks are introduced in the third thrust of this research. In order to evaluate those 3D synthetic datasets, a stochastic synthetic dataset quality assessment measure, named SSQA, is proposed that can fill the existing gap in the art to evaluate 3D synthetic dataset quality.

The automated algorithms, presented in this research, are able to promote the single-cell and population-level studies by combining super-resolution imaging with computational image advance. Furthermore, they also enable reconstruction and analysis of 3D biofilms. As conducted in the last thrust in this dissertation, biofilm reconstruction via geometrical model fitting and shape analysis of segmentation results can provide the cell biologists with statistics to further explore bacterial cell morphology and intrabiofilm mechanisms in future work.