Object- and spatial-level quantitative analysis of multispectral histopathology images for detection and characterization of cancer

Abstract

The main goal of this dissertation is the development and discussion of techniques for higher-level image analysis, i.e., object-level analysis, of breast cancer imagery. Established cytologic (cell) criteria can be contradictory, and even histologic (tissue) criteria (considered the gold standard for diagnosis) are subject to varied interpretation. There is thus a need to quantitatively define characteristics of breast cancer to bettor coordinate clinical care of women presenting breast masses. We propose here an approach for such quantitative analysis, Quantitative Object- and spatial Arrangement-Level Analysis (QOALA), using expert (pathologist) input to guide the classification process. The main contributions in this work are four-fold. First, quantitatively analyze the utility of multispectral imagery for classification and segmentation tasks in histopathology imagery. Second, we develop object-level segmentations for several histologic classes, as a quantitative object-level segmentation metric. Third, we extract a comprehensive set of both object- and spatial-level features which are used in a feature selection framework for classification of objects and imagery. Fourth, we extend the concepts of object-level features to higher-level image objects, analyze the utility of these high-level objects for image classification, and introduce the concept of a probabilistic graph-based model of imagery. Overall, QOALA yields very good object- and image-level classification performances. More specifically, the object-level features as implemented in QOALA are versatile and general enough to elicit important information from even imperfectly segmented objects. Additionally, the use of non-nuclear features, namely features of cytoplasm and stoma have good classification performance, often exceeding that of nuclei. Higher-level features display a potential to increase both object- and image-level classification performance.

Ähnliche Arbeiten