Notice:

• This repo is a set of tools extracted from a larger repository. https://github.com/ramon349/CUID-PFM#
• The split was done to provide only the wsi processing utilitiy. It will be refactored to improve utility.

Enviroment setup

• Install trident via editable install. See instructions here
• If running the tile processing. One should change the dataloaders used by trident to use multiprocessing_context='fork'.
• Install dependencies using reqs.txt
• The install takes a while. Future versions will try to reduce the number of dependencies.

conda create -n conch python=3.10
cd ./causal_path
python3 -m pip install -r ./reqs.txt 

• Install this repo as an editable module by doing

python3 -m pip install -e .

Dataset Preparation: Tile Level Extraction of Foundation Model Features

• All we need is a csv file with a column named wsi that specifies the path to an svs image.
• Fill out the details bellow
  • csv_path: has a column wsi with absolute path to svs images
  • output_dir: This is where the extracted features,logs, geojson, segmentation masks will be stored by our pipeline
  • token_dir: Path to the directory with your huggingface token. This is needed to access models
  • config_save_dir: We will fill this directory with config files for running each of the extraction steps
  • num_workers: DO NOT TOUCH
• The processing pipeline is rather top heavy. We will first create tissue segmentations for each slide image. This is the most time consuming step.
• Tile and feature extraction are rather quick. Note that tiles are stored as geojson. We do not save pngs of the tile regions themselves

Generate Processing Config Files

python3 -m causal_path.exp_setup.wsi_processing.make_proc_pipeline --csv_path /pathToYourFolders/data/csvs/colon_cancer_dev_set.csv \
--output_dir /pathToYourFolders/data/embeddings/crc_embeddings \
--wsi_dir /PathToDirectoryContainsWSI/Rawimages/ \
--num_workers 1 \
--token_dir "/pathToYourHGToken/hg_tok.json" \
--config_save_dir /pathToYourFolders/data/configs/embedding/crc

Run Tissue Segmentation

python3 -m causal_path.tiling --config_path /pathToYourFolders/data/configs/embedding/crc/segmentation.json

Extract Tiles/Patches

python3 -m causal_path.tiling --config_path /pathToYourFolders/data/configs/embedding/crc/patch.json

Extract Feature Embeddings

• Extract Embeddings. We only show conch_v1. But the code generates virchow_v2 and uni2 configurations as well. One can also expand to other models supported by trident.

python3 -m causal_path.tiling --config_path /pathToYourFolders/data/configs/embedding/crc/feat_extract/extract_conch_v1_feats.json