Morphological analysis of the appearance and quantity of cells and tissue architecture has been routinely used by pathologists to determine cancer’s histologic grade. In standard practice, however, the analysis is based on only a few morphological features, and is mainly done in a qualitative or semi-quantitative fashion. This makes the analysis subjective and less reproducible. In fact, a single piece of cancer tissue can contain a high level of heterogeneity – two cells of the same type can behave differently depending on their genetic makeup and their microenvironment. This simply means that there is still a vast amount of information hidden in cancer tissues, whose clinical value is not fully realised. With digital slide scanning technology becoming more available, it is now possible to obtain a whole-slide image of a histological section of high-quality and with high-resolution (multi-gigapixel images) in a digital format. Moreover, the volume of data is rapidly growing daily as slides are routinely scanned. This presents an opportunity to advance image analytical techniques and computational algorithms for quantitative analysis of tissue morphology (morphometrics). Morphometrics provides an accurate and reproducible means for the diagnosis and prognostication of cancers. This is the very first step towards effective treatment decision making and personalised medication. In the first part of this thesis, we focus on the development of automated algorithms for quantifying morphological features from histological sections. All the algorithms are specifically designed for histological sections stained by a standard haematoxylin and eosin (H&E) stain. Firstly, we proposed a discriminative dictionary learning based algorithm for learning the appearance of cells. The approach is applied to distinguish mitotic cells from non-mitotic cells in breast cancer histological images. Secondly, we develop a deep learning based framework that utilises a local context information for nuclear detection and classification. We demonstrate the effectiveness of this framework for detecting and classifying various types of cells in colorectal cancer. Lastly, we consider the intestinal gland segmentation problem in histological images of colorectal cancer. We formulate a segmentation algorithm that finds the best polygon matching the shape of each gland in a Bayesian framework. The method is capable of segmenting glands in both benign and malignant cases. In the second part of this thesis, we turn toward a quantitive analysis of tissue morphology of colorectal adenocarcinomas. This is to find potential morphological features that are clinically relevant. We extract morphological features, using the algorithms developed in the first part of this thesis. The results from statistical analyses show that a quantitative measure for aberrance of glandular shape is strongly associated with the histologic grade, and a quantitative measure related to the desmoplasia is associated with the tendency to develop metastasis.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:690511 |
| Date | January 2016 |
| Creators | Sirinukunwattana, Korsuk |
| Publisher | University of Warwick |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://wrap.warwick.ac.uk/80222/ |
Page generated in 0.0015 seconds