Bilkent University
Department of Computer Engineering


Superpixel-Based Segmentation of Histopathological Images


Mert Keser
MS Student
Computer Engineering Department
Bilkent University

Today, histopathological examination is the routinely used practice for cancer diagnosis and grading. In this practice, pathologists examine changes in the cell morphology and the distribution of the tissue components under a microscope and make decisions based on their visual interpretations. As this practice mainly depends on visual interpretation, it is open to subjectivity, which might lead to incorrect decisions. To reduce this subjectivity, it has been proposed to design automated decision systems. The first step of these systems is usually the segmentation of a tissue image into its homogeneous regions. In this work, we propose a new tissue segmentation algorithm that relies on the difference between the spatial distribution of the components in normal and cancerous regions. To this end, it first decomposes the image into super-pixels, each of which approximately corresponds to a tissue component, and labels them by the k-means algorithm. It then constructs a graph on these labeled super-pixels and defines a similarity measure for the adjacent ones by extracting graph-based features. It finally uses these features in a region-growing algorithm to segment the image into its normal and cancerous regions.


DATE: 21 Kasım, 2016, Monday @ 16:40