El-Shanawany, A.; (2010) The development of generative Bayesian models for classification of cell images. Masters thesis , UCL (University College London).

Preview

PDF 20314.pdf Download (2MB)

Abstract

A generative model for shape recognition of biological cells in images is developed. The model is designed for analysing high throughput screens, and is tested on a genome wide morphology screen. The genome wide morphology screen contains order of 104 images of fluorescently stained cells with order of 102 cells per image. It was generated using automated techniques through knockdown of almost all putative genes in Drosphila melanogaster. A major step in the analysis of such a dataset is to classify cells into distinct classes: both phenotypic classes and cell cycle classes. However, the quantity of data produced presents a major time bottleneck for human analysis. Human analysis is also known to be subjective and variable. The development of a generalisable computational analysis tool is an important challenge for the field. Previously cell morphology has been characterized by automated measurement of user-defined biological features, often specific to one dataset. These methods are surveyed and discussed. Here a more ambitious approach is pursued. A novel generalisable classification method, applicable to our images, is developed and implemented. The algorithm decomposes training images into constituent patches to build Bayesian models of cell classes. The model contains probability distributions which are learnt via the Expectation Maximization algorithm. This provides a mechanism for comparing the similarity of the appearance of cell phenotypes. The method is evaluated by comparison with results of Support Vector Machines at the task of performing binary classification. This work provides the basis for clustering large sets of cell images into biologically meaningful classes.

Download activity - last month

Export as JSONCSVXML

Download activity - last 12 months

Export as XMLCSVJSON

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item