zeynepozdemir / oeem Goto Github PK

Note: this code is expected to be ready at middle June.

Yi Li*, Yiduo Yu*, Yiwen Zou*, Tianqi Xiang, Xiaomeng Li, "Online Easy Example Mining for Weakly-supervised Gland Segmentation from Histology Images", MICCAI 2022 (Accepted). [paper]

This framework is designed for histology images, containing two stages. The first classification stage generates pseudo-masks for pathes. And the segmentation stage uses OEEM to mitigate the noise in pseudo-masks dynamically.

This code has been tested with Python 3.7, PyTorch 1.10.2, CUDA 11.3 mmseg 0.8.0 and mmcv 1.4.0 on Ubuntu 20.04.

Download pretrained models and dataset, then link to codes.

git clone https://github.com/XMed-Lab/OEEM.git
cd OEEM
ln -s [path of patches for cls] classification/glas
ln -s [path of patches for seg] segmentation/glas
ln -s [path of models for cls] classification/weights
ln -s [path of models for seg] segmentation/models

Install library dependencies

pip install -r requirements.txt

Install mmsegentation.

cd segmentation
pip install -U openmim
mim install mmcv-full==1.4.0
pip install -v -e .

Train classification model.

python classification/train.py -d [gpu device no.] -m [model_name]

Generate pseudo-mask (WSI size). The output will be in [model_name]_best_train_pseudo_mask folder.

python classification/prepare_seg_inputs.py -d [gpu device no.] -ckpt [best_model_name]

Split WSI pseudo-mask to patches for segmentation.

Train segmentation model.

Test segmentation model.

Merge patches and evaluation.

Results compared with WSSS for natural images:

This repository is released under MIT License (see LICENSE file for details).