TAD-Graph

TAD-Graph: Enhancing Whole Slide Image Analysis via Task-Aware Subgraph Disentanglement

Installation

conda create -n wsi python=3.9
conda activate wsi
sh INSTALL_TORCH.sh
pip install -e .

Dataset Preparation

We follow CLAM to divide each WSI into patches and extract features for each slide.

1. For patching:

cd src/tadgraph/preprocess
python create_patches_fp.py --task tcga_prad --seg --patch --stitch 

2. Extract patch embeddings:

cd src/tadgraph/embedder
CUDA_VISIBLE_DEVICES=6,7 python extract_features_fp.py --task cptac_brca --data_h5_dir extracted_mag20x_patch256 --model uni --batch_size 256

Usage

cd scripts

Please refers to the command in main.py.

For example:

CUDA_VISIBLE_DEVICES=6 python main.py --model_name tadgraph --config tad_graph_config.yaml \
--dataset tcga_brca --task her2 --split_file tcga_brca_her2_5fold_val0.2_test0.2_100_seed1  \
--feat_dir extracted_mag20x_patch256/vits_tcga_pancancer_dino_pt_patch_features/slide_graph --embed_size 384 --use_graph \
--lambda_sup 1 --lambda_info 0.5 --lambda_unif 0.5

Acknowledgement

We thank the awesome CLAM and HIPT codebase.