Bilkent University
Department of Computer Engineering
MS THESIS PRESENTATION
Local Object Patterns for Tissue Image Representation and Cancer Classification
Gülden Olgun
MSc Student
Computer Engineering Department
Bilkent UniversityHistopathological examination of a tissue is one of the routine practice for diagnosis and grading of the cancer. However, the result of this examination is subjective since it requires correct interpretation which is contingent on the pathologists' expertise and experience. In order to minimize the effect of the subjectivity, automated cancer diagnosis and grading systems which represent the image with quantitative features and use these features for classifying and grading the cancer are proposed. Mostly, textural and structural approaches are used for automated cancer diagnosis and grading systems. However, textural approaches are sensitive to the pixel-level noises whereas structural approaches represent the entire image with a single graph. Therefore, we propose a new algorithm which defines a new set of texture descriptors is called local object patterns for effective representation and classification of histopathological colon tissues stained with hematoxylin-and-eosin. Thus, we decompose the tissue image into histological components and introduce the local object patterns to model the components by using the idea of the local binary patterns. Though, our proposed algorithm defines local object patterns on the objects at the component-level is differentiated from the local binary patterns which defined at pixel-level. In order to define the textures, we used two approach for selecting the component objects. First approach, we call n-LOP approach, defines descriptors for each object, but, the other approach, we call GraphWalk, use a set of sub-graphs by walking over the image graph to define the descriptor. To this end, binary string is construct for components's local neighbors with different locality and describe spatial arrangements of the components for these neighborhood to extract the texture descriptors. From these texture descriptors, we create visual words for the bag-of-words representation to classify the images. Working with histopathological tissue images, our experiments shows that our two approaches give higher classification accuracies than the other approaches.
DATE: 30 July, 2013, Tuesday @ 10:00
PLACE: EA-409