Fig. 4: A hierarchical GNN-based model was built for predicting lymph node metastasis. | npj Precision Oncology

Fig. 4: A hierarchical GNN-based model was built for predicting lymph node metastasis.

From: Artificial intelligence-driven prediction of lymph node metastasis in T1 esophageal squamous cell carcinoma using whole slide images

Fig. 4

The workflow comprises: (1) Input WSIs are divided into 224 × 224 pixel patches (stride 112), followed by spatial domain preprocessing and graph construction. (2) ResNet-50 extracts patch features (2048 channels), reduced to 512 channels via dimensionality reduction, and fused through multimodal integration. (3) A hierarchical architecture with two-stage graph convolutional layers (GCNConv), feature compression, and global context aggregation modules learns spatial dependencies and semantic representations. (4) Global mean pooling and a softmax classifier perform binary classification, outputting invasion probabilities.

Back to article page