Computerized detection, segmentation and classification of digital pathology: case study in prostate cancer

Digital pathology refers to the use of scanning hardware and viewing software to digitize samples of stained pathological tissue excised from a patient. Image analysis algorithms can be employed to assist in analyzing these digital samples, increasing the speed and efficiency with which pathology samples are examined in the clinic. Traditionally these algorithms have focused on simple quantification (e.g. cell counting or stain enhancement), but the most recent developments have focused on developing quantitative disease signatures for different tissue types. In this dissertation, an image analysis framework for automated interpretation of histology samples is described using novel image descriptors and new classification techniques. This interpretation of samples has several advantages over the traditional method of manual analysis: (1) by using quantitative disease metrics, it can be applied in a standardized fashion across several institutions with perfect agreement; (2) advanced pattern recognition and machine intelligence algorithms such as supervised classification, intelligent training, and content-based image retrieval can be employed to add to the information used to make a decision regarding diagnosis and treatment; and (3) by providing such an in-depth analysis of tissue, we can make predictions regarding the potential outcome of patients with respect to specific treatment regiments. The overall goal of the framework is to reduce the burden on pathologists who must examine hundreds of thousands of tissue images every year, and to enhance the ability of clinicians to detect, diagnose, and treat disease. We apply our framework to a series of datasets with a focus on detection, segmentation, and classification of prostate cancer. Our data consists of over 100 patient biopsy samples stained with hematoxylin and eosin and digitized at 40x optical magnification. Ground truth for normal, diseased, and confounder tissue was manually applied by expert pathologists. We demonstrate the ability of our framework to perform the following: detect suspicious regions of tissue on whole biopsies; analyze those regions in detail using morphological, textural, and architectural characteristics to correctly classify each region; provide an intelligent method for training the classifier and identifying new tissue classes; and perform content-based image retrieval of images in the database.Ph.D.Includes bibliographical referencesIncludes vitaby Scott Doyl