Shape and Texture Recognition for Automated Analysis of Pathology Images

• Main Author: Snell, Violet
• Format: Dissertation
• Language: English
• Published: ProQuest Dissertations & Theses 01-01-2014
• Subjects: Computer science; Medical imaging
• ISBN: 1392687365; 9781392687369

This research project is concerned with automated analysis of microscopic images used in clinical pathology for diagnosing disease. Application of computer vision methods can improve the accuracy, reliability and availability of tests, reduce the associated costs and ultimately improve patient outcomes. Three different areas of pathology are covered: • identification of clustered nuclei and detection of chromosomal abnormalities in DAPI-stained samples,• diagnosis of auto-immune diseases from indirect immunofluorescence (IIP) images, and • detection of dividing nuclei in H&E stained histopathology sections. Despite the diversity of these application domains, the techniques used for their analysis are similar. For cluster identification in DARI images we focus on object shape and extend existing methods of shape analysis with novel measurements of the boundary profile which detect notches between overlapping nuclei in a cluster. For abnormality detection we focus on texture and develop a novel decision-tree dictionary for patch quantisation. We continue to focus on texture for IIP images, developing suitable isotropic measurements as well as exploring the connections between classification of individual cells and whole patient samples. Detection of dividing cells in tissue sections requires a combined assessment of shape, texture and colour in order to fully represent all relevant facets of the object. Here we develop a method for stain normalisation which efficiently compensates for batch variations in stain strength and proportions, followed by a full pipe-line of segmentation, feature extraction and classification, resolving issues of class imbalance implicit in detection of rare objects. We develop an efficient and effective segmentation method, which is free of weight parameters and adaptable for use in different imaging modalities. We explore a variety of classifier types and ensemble structures, and suggest promising directions of future development in the broad application area of pathology image analysis.