Using Hierarchically Connected Nodes and Multiple GNN Message Passing Steps to Increase the Contextual Information in Cell-Graph Classification

Graphs are useful in analysing histopathological images as they are able to represent neighbourhood interactions and spatial relationships. Typically graph nodes represent cells and the vertices are constructed by applying a nearest neighbor algorithm to cell's locations. When passing these graphs through one graph neural network (GNN) message passing step, each node can only utilise features from nodes within its immediate neighbourhood to make a classification. To overcome this, we introduce two levels of hierarchically connected nodes that we term "supernodes". These supernodes, used in conjunction with at least four GNN message passing steps, allow for cell node classifications to be influenced by a wider area, enabling the entire graph to learn tissue-level structures. The method is evaluated on a supervised task to classify individual cells as belonging to a specific tissue class. Results demonstrate that the inclusion of supernodes with multiple GNN message passing steps increases model accuracy.