Tumor Detection & Classification in Breast Cancer Histology Images using Deep Neural Networks

Type: MA thesis

Status: finished

Date: July 5, 2019 - December 31, 2019

Supervisors: Sulaiman Vesal, Andreas Maier, Tino Haderlein

Among females, breast cancer is one of the most frequently diagnosed cancers and the leading causes of cancer-related death both worldwide, and in more economically developed countries. Early diagnosis significantly increases treatment success, since the treatment is more difficult and uncertain when the disease is detected at advanced stages. For this purpose, proper analysis of histology images is essential. Histology is the study of the microanatomy of cells, tissues, and organs as seen through a microscope.

One of the most common type of Histology images used as the basis of contemporary cancer diagnosis for at least a century is Hematoxylin and eosin (H&E) stained breast histol- ogy microscopy images[4]. During this diagnosis procedure, trained specialists evaluate both overall and local tissue organization of the images. However, due to the large amount of data and the complexity of the images, this task becomes very time consuming and non-viable. Therefore, the development of software tools for automatic detection and diagnostic tools is a promising prospect in this field. This subject has been a rather active field of research, and thus, the automatic detection of breast cancer based on histology images is part of the ICIAR 2018 challenge on BreAst Cancer Histology (BACH) challenge. This challenge consists of two parts; classification and segmentation.

The aim of this thesis is to first design a classifier network, which can recognize types of breast cancer. Then, using another network, we will try to classify the WSIs and perform segmentation on the images. Afterwards, we want to investigate how weakly-supervised training can affect our results on both image-wise (first part) and pixel-wise labeled images (second part). For this purpose, we will start with reproducing the results of the winning paper, which is the state of the art. Then we try to build the rest on top of that.