README.md

Joint Fully Convolutional and Graph Convolutional Networks for Weakly-Supervised Segmentation of Pathology Images

Instructions

A trained checkpoint numbered 1000 is provided with 9 HER2 pathology images for use in inference.
This checkpoint is trained with 226 HER2 pathology images from a private dataset

To run inference:

python3 finaledgegcncopy.py --inference-path full_path_to/inference --checkpoint xxxx for example, with provided images and state dict, run like:
python3 finaledgegcncopy.py --inference-path full_path_to/inference --checkpoint 1000

To run Train:

python3 finaledgegcncopy.py

To resume Train:

python3 finaledgegcncopy.py --checkpoint xxxx

Flags and folders:

--train-path or ./train_process_files: a folder which the pipeline saves training visualization files to

--input-path or ./input_data: images used for inference or training. For our weakly supervised loss to work, the training images should be named as: AreaRatio_Uncertainty_.png.
For example, 0.4_0.05_.png means the target region occupies (40+/-5)% of the image.

--inference-path or full_path_to/inference: an argument that defines a folder for the pipeline to output inferenced mask to. Setting this argument will switch on inference mode. This argument must be used with --checkpoint.