• NumPy
• Pandas
• Scikit-learn
• Tqdm
• PyTorch
• PyTorch Geometric

First, download TCGA data. You need to cut whole-slide images (WSIs) into patches (e.g. size $256\times 256$ at $20\times$ magnification) , and then generate the instance-level features using an encoder network. The instance-level features of each WSI are used to generate the WSI bag. Each WSI bag is stored as a .pt file.

After that you need to generate an index file (in .pth format), which stores a python list, and each item of the list is a dict, which corresponds to one WSI information.

[{'slide': 'xxx.pt', 'shape': torch.Size([N, D]), 'target': 'y'}]

• slide: the file of WSI bag
• shape: the size of WSI bag
• target: the label string of WSI

Perform hierarchical global-to-local sampling strategy to build visual codebook and generate the sampled WSI bags.

$ python codebook_generation.py

Construct WSI graphs using the sampled WSI bags.

$ python graph_construction.py

$ CUDA_VISIBLE_DEVICES=0 python main_nsclc.py 

If you want to cite this work, please use the following BibTeX entry.

@InProceedings{Guan_2022_CVPR,
    author    = {Guan, Yonghang and Zhang, Jun and Tian, Kuan and Yang, Sen and Dong, Pei and Xiang, Jinxi and Yang, Wei and Huang, Junzhou and Zhang, Yuyao and Han, Xiao},
    title     = {Node-Aligned Graph Convolutional Network for Whole-Slide Image Representation and Classification},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {18813-18823}
}