The study of stem cells has received considerable attention in forming many different tissue types, and gives hope to many patients as it provides great potential for discovering treatments and cures to many diseases such as Parkinson's disease, schizophrenia, Alzheimer's disease, cancer, spinal cord injuries and diabetes. This study was concerned with developing algorithms that analyses microscope images of stem cells harvested from the bone marrow or dental pulp of a rabbit, expanded in the laboratory at the Tissue Engineering Center in Alexandria, Egypt, and then transplanted into subcutaneous pouches of the rabbit. The research aimed to detect automatically as soon as osteogenic differentiated stem cells were ready to be implanted in the defective parts, thereby avoiding the cells becoming damaged by bacterial infection. A further requirement was that the algorithms would not use traditional (chemical) markers which eventually lead to the sample being discarded as it dies after adding the marker. A total of 36 microscopy images were obtained from seven separate experiments each lasting over 10 days, and the clinicians visually classified 18 images as showing not-ready osteogenic differentiated stem cells and the remaining images showing a variety of cells ready for implantation. The ready cells typically appeared as a colony, or spread all over the image interconnecting together to form a layer. Initially, image pre-processing and feature extraction techniques were applied to the images in order to try and identify the developing cells, and a t-test was applied to the total cell area in each image in an attempt to separate the not-ready and ready images. While there was a significant difference between not-ready images and the ready images which showed the colony shaped characteristics, there was no significant difference between not-ready images and ready images with the spreading interconnecting layer shape, and so more sophisticated classification techniques were investigated. As the differentiated stem cells are effectively texture based images, each of the 36 images were divided into quadrants to give a total of 144 images to increase the image dataset. Several sets of texture parameters were derived from the grey-scale histogram statistics, Grey-Level Co-occurrence Matrix (GLCM), and Discrete Cosine Transform (DCT) spatial frequency components of the images. Some of these parameters were used with traditional classification techniques including cross-correlation, and Euclidean distance measures to try and classify the texture relative to the first image (not-ready) in each experiment and the other images (not-ready and ready) in the experiment. The success rate using cross-correlation was 70%, and 68% for the Euclidean distance approach. Secondly, intelligent classification techniques using Artificial Neural Networks (ANN) were considered, using the various texture parameters as inputs to a feed-forward 1-hidden layer MLP using Back-propagation of Errors for training. The ANN approach gave the better results, with 77% using the grey-scale histogram statistics, 73% for GLCM, and 92% for the DCT with 70 spatial frequency components. It was observed for each of the experiments that images became classified as ready for implantation after approximately 10 days, and then remained ready for the rest of the experiment.
Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:577200 |
Date | January 2013 |
Creators | abdelgawadbirry, R. |
Publisher | University of Salford |
Source Sets | Ethos UK |
Detected Language | English |
Type | Electronic Thesis or Dissertation |
Source | http://usir.salford.ac.uk/29225/ |
Page generated in 0.0024 seconds