Time-lapse image analysis for understanding mycobacterial growth and persistence

• Main Author: Hu, Yin
• Other Authors: Tang, Hongying
• Published: University of Surrey 2017
• Subjects: 004
• Online Access: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.731155

As one of the world’s most devastating diseases of mankind, tuberculosis is a global health crisis. Despite extensive research into the disease spanning more than a century, it remains the number one killer due to a single causative and infectious agent, Mycobacterium tuberculosis, which is one of the human pathogens. The bacterium is able to persist as a long-term infection, known as latent tuberculosis. One characteristic of persister cells is that they are phenotypically tolerant to the action of antibiotics; a trait which has important implications in tuberculosis chemotherapy. Observing persistence is an important element of the latent tuberculosis study. To gain insight into persistence of tuberculosis, confocal microscopy is used to capture the cell growth and division behaviour in a microfluidic device so that a large amount of time-lapse image data are collected. It is very challenging for human observers to grasp them directly. In this work, we aim to develop a system that is able to track and analyse the cell growth patterns automatically. Specially, a major task is to observe any unusual behaviour, such as persistence. We first presented an approach to extract cell properties evolving between consecutive frames by feeding cell segmentation and tracking results from one frame to the next. Each individual cell is obtained by integrating the Distance Regularised Level Set Evolution model with cell septum and membrane function. It was then tracked by minimising a single cell trajectory energy function along time-lapse series. Our experiments showed that cell growth and division can be measured automatically by applying this scheme. Comparing with other existing algorithms, our results showed the efficiency of the approach when testing on different datasets. The proposed approach has demonstrated great potential for large scale bacterial cell growth analysis. We further investigated the deep convolutional neural networks with a hierarchical visual tracking approach. We demonstrated that this approach can robustly segment and track individual cells from different microscopy image types, such as phase-contrast and bright-field microscopy images. Comparing with previous methods, the convolutional neural networks have significantly improved accuracy in cell segmentation thus minimising manual correction effort. We also outlined several rules for designing and optimising deep convolutional neural networks for this study. We believed that deep convolutional neural network approach is robust for segmenting and tracking various strains of bacteria cells. With above analysis, we obtained a large number of cell growth features over several generations. We discovered that the loss of phenotypic inheritance causes increased frequency of persisters. We also illustrated that cell growth and division was most consistent with the adder model in a single generation. These novel observations can be accounted for the generation and maintenance of phenotypic variation and provide potential new targets for the development of novel therapeutic strategies that address persistence in bacterial infections.