Department of Computer Engineering
CS 590 SEMINAR
Breast Cancer Detection and Recognition on Whole Slide Histopathology Images with Deep Convolutional Networks
Computer Engineering Department
Breast cancer is the most prevalent form of cancers among women. Medical image processing might have a huge contribution to experts for analysis of histopathology images by improving interpretation or indicating candidate disease locations. Although hand-crafted features are widely shown to be effective for discrimination of easier problems such as classification between healthy and invasive cancerous regions, it is not clear how much information is captured by those features when it comes to more challenging and less distinctive problems which sometimes experts might not agree upon a single diagnosis. In this study, we propose a complete system by using advanced deep learning techniques, by imitating actions of human pathologists on whole-slide pathology images (WSI) to diagnose breast cancer. First, we train multiple Fully Convolutional Networks (FCNs) for saliency detection on WSI from pathologists actions for efficiency of later steps. Then, one deeper CNN is trained for identification of several diagnostic classes to be run on salient regions detected by FCNs. Those networks are visualized for better understanding and evaluated comprehensively by comparing them to state-of-art methods and human performance.
DATE: 21 March, 2016, Monday @ 16:10