Bilkent University
Department of Computer Engineering


Weakly Supervised Region of Interest Classification in Breast Histopathology Images


Bulut Aygüneş
MS Student
Computer Engineering Department
Bilkent University

Whole slide digital imaging technology has been replacing the traditional pathology practices involving the examination of biopsy whole slides under the microscope. Being able to digitize whole slides scanned at very high resolutions aroused the interest of researchers for developing automated tools to help pathologists with the diagnostic process by using deep learning techniques, which are suitable for obtaining features to represent complex structures in breast biopsy images. However, since whole slides are very large, most of the related work focuses on smaller, diagnostically relevant regions of interest (ROI). These ROIs usually have a fixed size and are chosen among the diagnostic categories that are easier to distinguish between, which is not a realistic scenario. In our work, we focus on the classification of variable-sized ROIs with diagnostic labels from different sub-categories of cancer that are harder to distinguish, which is a more realistic and clinically important problem. We model the problem as a weakly supervised classification problem, considering that the diagnosis of an ROI is known, but how much different patches extracted from the ROI contribute to the given diagnosis is unknown. We experimented with a weakly supervised object detection model named Weakly Supervised Deep Detection Network (WSDDN) and compare it with the current baselines.


DATE: 06 May 2019, Monday, CS590 presentations begin at @ 15:40