Self-Supervised Representation Learning using Visual Field Expansion on Digital Pathology

The examination of histopathology images is considered to be the gold\nstandard for the diagnosis and stratification of cancer patients. A key\nchallenge in the analysis of such images is their size, which can run into the\ngigapixels and can require tedious screening by clinicians. With the recent\nadvances in computational medicine, automatic tools have been proposed to\nassist clinicians in their everyday practice. Such tools typically process\nthese large images by slicing them into tiles that can then be encoded and\nutilized for different clinical models. In this study, we propose a novel\ngenerative framework that can learn powerful representations for such tiles by\nlearning to plausibly expand their visual field. In particular, we developed a\nprogressively grown generative model with the objective of visual field\nexpansion. Thus trained, our model learns to generate different tissue types\nwith fine details, while simultaneously learning powerful representations that\ncan be used for different clinical endpoints, all in a self-supervised way. To\nevaluate the performance of our model, we conducted classification experiments\non CAMELYON17 and CRC benchmark datasets, comparing favorably to other\nself-supervised and pre-trained strategies that are commonly used in digital\npathology. Our code is available at https://github.com/jcboyd/cdpath21-gan.\n