Deep learning based histopathology image analysis for cancer diagnosis and treatment

Histopathology plays a vital role in cancer diagnosis, prognosis, and treatment decisions. The whole slide imaging technique that captures the entire slide as a digital image (whole slide image, WSI) allows the pathologists to view the slides digitally as opposed to what was traditionally viewed under a microscope. With the development of computational power and image analysis algorithms, computational methods have been developed for the quantitative and objective analyses of histopathology images, which can reduce the intensive labor and improve the efficiency for pathologists compared with manual examinations. In this dissertation, we focus on deep learning based solutions in histopathology image analysis for cancer diagnosis and treatment, specifically, nuclei segmentation for cancer diagnosis, and gene mutation and pathway activity prediction for cancer treatment. Nuclei segmentation is a critical step in the automatic analysis of histopathology images. We focus on two directions to tackle the problems in deep learning based nuclei segmentation task. One is the annotation-efficient algorithms. As the fully-supervised learning of deep neural networks requires a large amount of training data, a weakly supervised nuclei segmentation framework based on a portion of nuclear locations is proposed to alleviate the annotation effort of pathologists. This method achieves comparable performance as the fully-supervised methods and about 60$imes$ speed-up (10\% points) in the annotation time. The other direction is to improve the instance segmentation performance. The networks and cross entropy loss in current deep learning-based segmentation methods originate from image classification tasks and have drawbacks for segmentation. Therefore, we propose a full resolution convolutional neural network (FullNet) and a variance constrained cross entropy (varCE) loss to improve the fully supervised segmentation performance. Except for the cell-level heterogeneity that is routinely used for cancer diagnosis, it remains unclear for many cancers that how tissue structures in histopathology slides are related to genomic features like gene alterations and expression patterns. We develop a deep learning model to predict the genetic mutations and biological pathway activities directly from histopathology slides in breast cancer. The weight maps of tumor tiles are visualized to understand the decision-making process of deep learning models. Our results provide new insights into the association between pathological image features, molecular outcomes and targeted therapies for breast cancer patients.Ph.D.Includes bibliographical reference