Fig. 3: Visualizing spatial features and heterogeneous message passing in deep GNN layers.

A The sequence of images includes: (1) a patch image from a WSI depicting a cellular community; (2) a cell graph constructed based on cellular coordinates; (3) a cell graph with randomly altered cellular coordinates; and (4) a cell graph without cellular coordinates. B Depiction of node feature transformations at each layer of GNN processing, visualized using t-SNE. C Similarity matrices of cell identity matrices across seven cell types, illustrating changes through GNN layers, were computed using cosine similarity.