GrEp

GrEp is a fast and efficient graph-based post-processing workflow for the classification of epithelial cells into normal and malignant as depicted in the Figure below.

Setup GrEp Environment

#1. Create conda environment
conda env create -f environment.yml

#2. Activate GrEp environment
conda activate GrEp

#3. Install pytorch with pip
python 3.10 -m pip install torch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 --index-url https://download.pytorch.org/whl/cu118

Model Weights

To run the model, both models for node embedding (ResNet) and Node classification (GCN) must be downloaded:

1. Download ResNet weights

2. Download GCN weights

Tile inference

The model expects:

1. tile images such as '.png' or '.tif' at 0.5 MMP
2. Epithelial centroids saved as dict in '.mat' file with an ['inst_centroid'] entry

The model returns:

1. '.mat' files with epithelial centroids coordinates and class prediction (0 for normal, 1 for malignant)
2. if save_overlays is set to "True", the overlay predictions will be saved as well, making the model slower

To run the pipeline on a folder containing tiles:

srun python3.10 run_tile_inference.py \
--tiles_path 'Path_to_tiles_folder' \
--img_mod 'tif' \
--ep_centers_path 'Path_to_epithelial_nuclei_mat_folder' \
--resnet_path 'Path_to_resnet_weight/Resnet18_pretrained.pt' \
--gcn_path 'Path_to_GCN_weight/GCN.pt' \
--save_path 'Path_to_save_predictions'
--save_overlays "False"

WSI inference

to come !