Automated Analysis of Cellular Signalling Parameters based on Images and Videos of Fluorescence Microscopy

Traore, David Daouda (2020) Automated Analysis of Cellular Signalling Parameters based on Images and Videos of Fluorescence Microscopy. Doctoral thesis, University of Buckingham.

Preview

Text 1301722 Automated Analysis of Cellular.pdf Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (68MB) | Preview

Abstract

Rapid changes in computer vision technologies have enabled automatic perspectives for more disciplines that tend to need heavy intervention from human experts. Computational biology, which is the application field of this thesis, is one of those disciplines where computer technologies (software and hardware systems) are applied on cell biology research, drug discovery, and disease diagnosis. The research conducted in this thesis is primarily concerned with automating the analysis of calcium imaging data obtained by two-dimensional (2D) fluorescence microscopy (FM) over living cells. The thesis also presents a theoretical and empirical analysis of the state-of-the-art object detection techniques used in Region based Convolutional Neural Network (R-CNN) and proposes a new R-CNN scheme tailored for cellular object detection in FM data. The analysed images are either individual greyscale images or image stacks of cardiac myocytes stained with DNA markers or calcium indicators. The cells are special cardiac cells found in lung veins, called Pulmonary Vein sleeve Cells (PVC) and Neo-natal Rat Ventricular Myocytes (NRVM) extracted from heart muscle. PVC stains enable the analysis of Calcium signals effect on heart physiology, whereas NRVM images allow autophagy process measuring through accurate cell counting. In the thesis, we demonstrated that automated hotspot detection can be achieved with 79.75% of precision by a two-level segmentation procedure combining thresholding and statistical filtering of cellular regions. We also showed that photobleaching can be corrected by exponential curve fitting and baseline adjustment from normalised calcium traces with respectively a mean square error (MSE) varying between [0.09013, 6.41796] and an overall accuracy of 78.75% for a real-life dataset. Finally, based on the investigation carried over state-of-the-art object detection techniques developed in the past such as the Edge Boxes, the Selective Search, the Objectness Measure, and the Sliding Window paradigm, we demonstrated that a customised R-CNN framework based on a data-driven proposal box generation outperforms with less sampling rate (61 proposals/image) and more ground truth coverage ratio (GTCR of 99.40%).

Actions (login required)

View Item