Graph Neural Network-Based Multi-Scale Whole Slide Image Fusion for pT Staging of Muscle-Invasive Bladder Cancer.

Accurate primary tumor (pT) staging in muscle-invasive bladder cancer (MIBC) is crucial for treatment and prognosis. Current methods require time-consuming, labor-intensive microscopic evaluation by pathologists, with inherent interobserver variability. There is a need for AI-driven automated diagnosis of whole-slide images (WSI) to improve diagnostic efficiency while maintaining accuracy in pT staging. We obtained 281 H&E-stained WSI samples from the TCGA dataset for developing and validating a graph neural network (GNN)-based diagnostic model, and 83 additional samples from a hospital for external validation. The GNN method integrated multi-scale WSI data, evaluated using areas under the curve (AUC), accuracy, sensitivity, and specificity. A multi-scale attention mechanism was added to enhance model interpretability by capturing pT staging infiltration patterns. Diagnostic results were compared with those of three pathologists of varying expertise. We developed the multi-scale WSI-integrated GNN model for histopathological staging (T2/T3/T4) of MIBC. The model demonstrated excellent performance on external validation, achieving an AUC of 0.911 and an accuracy of 0.905. Interpretability analysis revealed distinct infiltration patterns for each T-stage, while diagnostic comparisons against both ground truth and three independent pathologists showed strong agreement, with a Cohen's kappa coefficient exceeding 0.876. The model developed based on graph neural network methods can integrate multi-scale information from whole-slide tissue images, allowing it to capture key infiltration patterns for muscle-invasive bladder cancer pT staging. This enables precise pT staging and visualizes the multi-scale tumor infiltration regions through attention scores, with accuracy showing strong consistency with expert pathologists.