Histopathology DatasetGAN: Synthesizing Large-Resolution Histopathology Datasets

Abstract

Deep learning-based methods have powered recent advancements in medical image segmentation, accelerating the field past previous statistical and Machine Learning-based methods [1]. This, however, has simultaneously created a need for large quantities of labeled data, which is difficult in domains such as medical imaging where labeling is expensive and requires expert knowledge. Semi-supervised learning (SSL) addresses these limitations by augmenting labeled data with large quantities of more widely available unlabeled data. Existing semi-supervised frameworks based on pseudo-labeling [2] or contrastive methods [3], however, struggle to scale to the high resolution of medical image datasets. In this work, we propose the Histopathology DatasetGAN (HDGAN) framework, an extension of the DatasetGAN framework for image generation and segmentation that scales well to large-resolution histopathology images. We make several adaptations on the original framework, including updating the generative backbone, selectively extracting latent features from the generator, and switching to memory-mapped arrays. These changes reduce the memory consumption of the framework, improving its applicability to medical imaging domains.