Capturing cellular topology in multi-gigapixel pathology images

In computational pathology, multi-gigapixel whole slide images (WSIs) are typically divided into small patches because of their extremely large size and memory requirements. However, following this strategy, one risks losing visual context which is very important in the development of machine learning models aimed at diagnostic and prognostic assessment of WSIs. In this paper, we propose a novel graph convolutional neural network based model (called Slide Graph) which overcomes these limitations by building a graph representation of the cellular architecture in an entire WSI in a bottom-up manner. We evaluate Slide Graph for prediction of the status of human epidermal growth factor receptor 2 (HER2) and progesterone receptor (PR) expression from WSIs of HE stained tissue slides of breast cancer. We demonstrate that the proposed model outperforms previous state-of-the-art methods and is more computationally efficient. The proposed paradigm of WSI-level graphs can potentially be applied to other problems in computational pathology as well.