Graph deep learning on whole slide image predicts the context-aware prognostic pathological features of renal cell carcinoma

Dependencies

• To install the dependencies for this project, see the "requirements.yaml"
• Tested on Nvidia TESLA V100 x 2 with CUDA 11.1

Step 1: Processing whole slide image (WSI) into superpatch-graph

What is the superpatch-graph?

• Superpatch-graph is the compressed representation of whole slide image into graph structure in memory efficient manner.
• Run the ./Superpatch_network_construction/supernode_generation.py
  • Users can simply run the above script with pre-defined sample data
  • Or, users can use your own whole slide image by setting the "--graphdir"
  • Output files
  • Compressed network as ".pt"
  • Node position information in "_node_location_list.csv"
  • Superpatch aggregated dictionary in "_artifact_sophis_final.csv"

Step 2: Training TEA-graph using superpatch-graph

• Users can predict the prognosis of entire host with tumor environment-associated context analysis using deep graph learning (TEA-graph)
• Run the ./main.py with appropriate hyperparameters
  • Users can simply run the above script with pre-defined parameters and datasets
  • Or, users can use their own dataset preprocessed by "supernode_generation" script

Step 3: Visualization of IG (Integrated gradients) value on WSI

• Users can visualize the IG value which is highly correlated with risk value of each region in WSI
• Also, we provide subgraph-level contextual pathological feature extraction
• Run the ./IG_attention_feature_cal_main.py with same parameters you used for training your own TEA-graph model
  • Users must define the trained_parameters as "--load_state_dict"
  • "IG_analysis" directory is created inside the directory you choose as the "--load_state_dict"
  • Subfolder for each patient is created inside the "IG_analysis"
  • "attention_value.npy" indicates the edge-level attention value
  • "Node_IG_sophis.npy" indicates the node-level IG value
  • "whole_feature.npy" is the trained contextual feature through GNN
  • "_WSI_Image_mask_IG_new.gif" is the heatmap of IG value on WSI
  • "_WSI_graph_wo_IG.jpeg" is the superpatch-graph
  • "_WSI_graph_w_IG.jpeg" is the IG value colored superpatch-graph
  • "IG_again" directory is also created inside each patient's folder
  • "_IG_TME_subgraph.csv" indicates the each IG group's subgraph

Step 4: Biomarker discovery

• Users can extract the contextual biomarker using the calculated IG values and extracted feature at the previous step
• Run the ./Context_marker_discovery_main.py with approprate directory path
• Users can obtain the several candidate pathology images with visualized graph for contextual biomarker

Acknowledgments

• http://github.com/mahmoodlab/Patch-GCN
• http://github.com/lukemelas/EfficientNet-PyTorch
• http://github.com/pyg-team/pytorch_geometric

BiNEL (http://binel.snu.ac.kr) - This code is made available under the MIT License and is available for non-commercial academic purposes