Enhancing cross-institute generalisation of GNNs in histopathology through multiple embedding graph augmentation (MEGA)
Campbell J., Vanea C., Salumae L., Meir K., Hochner-Celnikier D., Hochner H., Laisk L., Ernst L., Lindgren C., Xie W., Nellaker C.
Many recent methods for the analysis of histology whole slide images (WSIs) have used graph neural networks (GNNs) to aggregate visual information over a large image resolution. However, domain shift is a significant challenge in computational histopathology, due to differences in WSI appearance between institutes, and the effect of these differences on training GNNs has not been explored. In this work, we present the Multiple Embedding Graph Augmentation (MEGA) strategy to improve the cross-institute generalisation of GNNs in histology. We show that by introducing image augmentation and normalisation to the node features used to train a GNN, we can train a model that is robust to domain shift without additional labels or further training of the feature extractor. We compare MEGA to noise-based regularisation and demonstrate its effectiveness in a node classification tissue prediction task in placenta histology.