Weakly-Supervised Semantic Segmentation for Histopathology Images Based on Dataset Synthesis and Feature Consistency Constraint

Tissue segmentation is a key part of the digital pathology workflow, allowing for the automatic extraction of regions of interest in an image. However, as with most deep learning methods, a significant amount of quality labeled data is required to train the models. This challenge is amplified by the high level of expertise required to annotate the data and the unique characteristics of histopathology images. To address these challenges, this thesis investigates the use of weakly-supervised segmentation approaches as a means to reduce the burden of annotation and thereby increase accessibility to tissue segmentation. Traditional weakly-supervised segmentation approaches can suffer from poor activation as a result of the model favouring the most discriminative features for the classification task. While Class Activation Map (CAM) based approaches often suffer from the problem of partial activation, we propose a novel approach to assist the model in creating more complete activation maps. We apply a pseudo-supervised contrastive loss (PSCL) which improves the activation and convergence of these models. We demonstrate significant increases in performance across domain-standard datasets and achieve particularly strong performance on the BCSS-WSSS dataset with respect to the state-of-the-art. v vi