Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.722777
Title: High-throughput image analysis of zebrafish models of Parkinson's disease
Author: Dong, Bo
ISNI:       0000 0004 6421 3427
Awarding Body: University of Sheffield
Current Institution: University of Sheffield
Date of Award: 2017
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
Light microscopy can be used to advance our understanding of the molecular and cellular biology related to human health and diseases. As a powerful new vertebrate model, zebrafish have been used in various research areas, particularly in cancer and Parkinson's disease research. Large-scale data extraction from microscopy is highly attractive because it enables unbiased multivariate analysis that could lead to systems medicine approaches. To obtain useful information from large-scale data, high-throughput image analysis methods and applications are desperately required. In this thesis, we have explored methods and developed applications for highthroughput light microscopy zebrafish image analysis, including addressing the key problems related to three-dimensional (3D) deconvolution/deblurring, robust feature detection and description, and object counting. In biological image analysis, dealing with out-of-focus light noise, low image quality, large-scale dataset, illumination, overlapping, occlusion and insufficient prior knowledge remains challenging. Methods to address the following problems have been presented in this thesis. The low image quality of fluorescence microscopy images is addressed in Chapter 3. Owing to the limitations of light microscopes, the whole imaging process can be considered as a convolution between the object and the point spread function. Out-of-focus light and noise cause loss of detail in captured images. Deconvolution is an ill-posed inverse problem; thus, regularization methods are required to obtain better results. A maximum a posterior approach with a novel regularisation strategy to remove out-of-focus blur in 3D fluorescence microscopy images is introduced in Chapter 3. The second problem is dyed dopaminergic neurone detection in zebrafish RGB optical sections (Chapter 4). Owing to the large-scale of the image, low image quality, irregular appearance of neurones and touching situations, manually counting individual neurones via the microscope can be labour-intensive, time-consuming, subjective, and error-prone. To solve this problem, this thesis explores different methods to detect individual neurones in 3D zebrafish RGB images, including using detectors with many different handcrafted features and features learned automatically from deep learning architectures. An additional class-imbalanced problem is discovered during the experiments involving the training of patch-based deep learning techniques using a large-scale dataset that contains a limited number of positive samples. To solve this problem, a dynamic cascade framework with deep learning architectures is designed. The last problem is cell counting in fluorescent microscopy images (Chapter 5). Detecting individual cells in two-dimensional (2D) fluorescence microscopy images is difficult owing to overlap. Rather than counting-by-detection methods, a countingby- regression method with an interactive interface for cell counting in a fluorescent image is proposed. Sparse Bayesian Poisson regression based on a Relevance Vector Machine framework is also proposed. The proposed framework enables accurate counting of a discrete number of cells and leads to much sparser models, which results in faster performance and maintains a comparable generalisation error.
Supervisor: Frangi, Alejandro ; Shao, Ling Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.722777  DOI: Not available
Share: