Fig. 1: Illustration of the comprehensive pipeline for HCC prognosis prediction and biomarker discovery.

A Overview of cell segmentation and classification, the construction of cellular networks for intercellular communication, and the subsequent prediction of patient risk scores based on communication between cell communities, which are then used for downstream tasks. B Initial full-cell segmentation and classification on WSIs are performed by junior pathologists and refined by senior experts. Cellular communities are then selected based on the diversity of cell types. The cell graphs are constructed by combining the centroids of cells with the minimum spanning tree (MST) algorithm. A GNN with a heterogeneous message-passing function and an interpretable gated attention module extracts local features from cell graphs constructed based on cellular communities. Then, a Transformer encoder processes interactions between cellular communities. Finally, a fully connected layer calculates the risk score for HCC recurrence, which is utilized for Kaplan-Meier curve analysis and time-dependent receiver operating characteristic (ROC) analysis. Spatial biomarker discovery is conducted through the interpretable gated attention module along with cell graph and cellular community distribution. C In the innovative HeteroMessage Graph Neural Network module, the local cell graph is used as input. The heterogeneous message passing function’s identity matrix explicitly represents cell types to facilitate communication between cells. The interpretable gated attention module quantifies the attention scores of each cell.