Towards The Identification Of Histology Based Subtypes In Prostate Cancer

With the advent of deep neural networks (DNNs), methods of semantic segmentation in histology have improved to a degree that it now possible to analyse morphological features that are not easily accessible to human interpretation. Such features may be used to stratify tissue subtypes in health and disease. A major obstacle is that DNNs require a large amount of data to achieve high performance and generalise to new patients. Such a requirement is unsuitable for exploratory investigations that only have access to small patient cohorts. In this work, we demonstrate how variational autoencoders and generative adversarial networks can be combined to generate realistic histology images suitable for training semantic segmentation models, resulting in a novel data-augmentation method for histology. Subsequently, we analyse if such models can be used to identify subpopulations of prostate glands with different molecular profiles. If successful, this development will ultimately lead to the discovery of novel disease relevant histology-based subtypes. We demonstrate that morphological features derived from the H&E images alone are sufficient to identify expression of a clinical biomarker in prostate glands.