Local object patterns for tissue image representation and cancer classification

Abstract

Histopathological examination of a tissue is the routine practice for diagnosis and grading of cancer. However, this examination is subjective since it requires visual interpretation of a pathologist, which mainly depends on his/her experience and expertise. In order to minimize the subjectivity level, it has been proposed to use automated cancer diagnosis and grading systems that represent a tissue image with quantitative features and use these features for classifying and grading the tissue. In this thesis, we present a new approach for effective representation and classification of histopathological tissue images. In this approach, we propose to decompose a tissue image into its histological components and introduce a set of new texture descriptors, which we call local object patterns, on these components to model their composition within a tissue. We define these descriptors using the idea of local binary patterns. However, we define our local object pattern descriptors at the component-level to quantify a component, as opposed to pixel-level local binary patterns, which quantify a pixel by constructing a binary string based on relative intensities of its neighbors. To this end, we specify neighborhoods with different locality ranges and encode spatial arrangements of the components within the specified local neighborhoods by generating strings. We then extract our texture descriptors from these strings to characterize histological components and construct the bag-of-words representation of an image from the characterized components. In this thesis, we use two approaches for the selection of the components: The first approach uses all components to construct a bag-ofwords representation whereas the second one uses graph walking to select multiple subsets of the components and constructs multiple bag-of-words representations from these subsets. Working with microscopic images of histopathological colon tissues, our experiments show that the proposed component-level texture descriptors lead to higher classification accuracies than the previous textural approaches.