Department of Computer Engineering
CS 590 SEMINAR
Unsupervised Deep Binary Representation Learning for Breast Histopathology Images
Computer Engineering Department
The interest in automatic image analysis of whole-slide histopathology images has been increasing in recent years, prompted by the introduction of whole-slide scanners into pathology labs and the acceptance of digital slides as a primary diagnostic tool. This area should have benefited from the recent advancements in computer vision utilizing deep learning methods since the complex patterns occurring in histopathology images prevented traditional models with hand-crafted features from achieving human-level performance. However, image sizes reaching100,000 x 100,000 pixels and the lack of large annotated datasets due to their high costs are making the task challenging. Therefore, unsupervised deep learning in which annotated data is not required has become a primary interest in the medical imaging community.
Learning to hash is regarded as an efficient approach for image retrieval. Recently, deep learning is adopted for image hashing. While supervised deep learning has been proved to produce leading performance compared with the non-deep methods, unsupervised deep hashing techniques have not achieved the same success. In our research, we work on applying unsupervised deep learning techniques to learn binary representations of breast histopathology images. More specifically, Generative Adversarial Networks and Autoencoder models that are modified to learn binary representations are trained on histopathology image patches. Success in learning robust representations in patch level would lead to higher performance in weakly-supervised classification and retrieval tasks in whole-slide level which can be utilized in computer-aided diagnosis tools.
DATE: 22 April 2019, Monday, CS590 presentations begin at @ 15:40