Olivér Atanaszov: Efficient Segmentation of Histopathological Images

BCCN Berlin / Technische Universität Berlin

Abstract

Histopathological examination is essential for the timely diagnosis of various diseases. Digital pathology allows the development of computer-aided diagnosis systems using image processing and machine learning algorithms. In particular, Deep Learning is highly capable of learning complex visual recognition tasks from large amounts of data. These systems have the potential to assist specialists in routine diagnostics and treatment development as well as the pharmaceutical industry by accelerating drug research. However, learning complex tasks requires large models which in turn makes these methods data-hungry. However, annotation data annotated data in medical fields is scarce. Thus, a common bottleneck of computer vision applications, - especially in such specialized domains as histopathology - is to gather a sufficient amount of human-labeled data. This thesis focuses on the semantic segmentation of histopathological images that is a quintessential problem in the field. More specifically, we suggest methods that try to alleviate the problem of scarcely available data in the context of segmentation. First, we evaluate a number of methods that rely on inductive biases in order to synthetically extend the training data. Afterwards, we propose a simple yet powerful framework that allows us to make use of large amounts of unlabeled data in a semi-supervised setup. Furthermore, we propose a method to handle sparsely annotated data. We demonstrate that these generic techniques can significantly improve the generalization performance of our models without acquiring further annotation. The proposed solutions in this work can be adopted to Deep Learning applications in most medical fields and have the potential to facilitate the adoption of Deep Learning into clinical routine as well as pharmaceutical research.

 

Additional Information

Master Thesis Defense

 

Organized by

Prof. Klaus-Robert Müller   & Prof. Frederick Klauschen   / Margret Franke