Color graph representation for structural analysis of tissue images

Abstract

Computer aided image analysis tools are becoming increasingly important in automated cancer diagnosis and grading. They have the potential of assisting pathologists in histopathological examination of tissues, which may lead to a considerable amount of subjectivity. These analysis tools help reduce the subjectivity, providing quantitative information about tissues. In literature, it has been proposed to implement such computational tools using different methods that represent a tissue with different set of image features. One of the most commonly used methods is the structural method that represents a tissue quantifying the spatial relationship of its components. Although previous structural methods lead to promising results for different tissue types, they only use the spatial relations of nuclear tissue components without considering the existence of different components in a tissue. However, additional information that could be obtained from other components of the tissue has an importance in better representing the tissue, and thus, in making more reliable decisions. This thesis introduces a novel structural method to quantify histopathological images for automated cancer diagnosis and grading. Unlike the previous structural methods, it proposes to represent a tissue considering the spatial distribution of different tissue components. To this end, it constructs a graph on multiple tissue components and colors its edges depending on the component types of their end points. Subsequently, a new set of structural features is extracted from these “color graphs” and used in the classification of tissues. Experiments conducted on 3236 photomicrographs of colon tissues that are taken from 258 different patients demonstrate that the color graph approach leads to 94.89 percent training accuracy and 88.63 percent test accuracy. Our experiments also show that the introduction of color edges to represent the spatial relationship of different tissue components and the use of graph features defined on these color edges significantly improve the classification accuracy of the previous structural methods.