[IPMI2023]Human-machine Interactive Tissue Prototype Learning for Label-efficient Histopathology Image Segmentation

Introduction

This repository is an official implementation of the IPMI 2023 paper Human-machine Interactive Tissue Prototype Learning for Label-efficient Histopathology Image Segmentation.

Datasets

CAMELYON16

Breast Cancer Semantic Segmentation (BCSS)

Dependencies

conda create -n Proto2Seg python=3.8
conda activate Proto2Seg
pip install -r requirements.txt

Contrastive Learning based Encoder Training

Training

cd ./contrastive_pretrain
python train.py --config [path/to/config]

Prototype Identification based on Clustering

cd ./prototype_dict_building_and_coarse_segmentation
python cluster.py --config  [path/to/config]

Coarse Segmentation Prediction

cd ./prototype_dict_building_and_coarse_segmentation
python coarse_seg_cluster_query.py --config  [path/to/config] --n 5

Refinement

Training

cd ./refinement
python -m torch.distributed.launch --nproc_per_node 8 train_seg.py --config [path/to/config]

Inference

cd ./refinement
python test.py --dir [path/to/log] --dataset-name [dataset]

Citation

@article{pan2022human,
  title={Human-machine Interactive Tissue Prototype Learning for Label-efficient Histopathology Image Segmentation},
  author={Pan, Wentao and Yan, Jiangpeng and Chen, Hanbo and Yang, Jiawei and Xu, Zhe and Li, Xiu and Yao, Jianhua},
  booktitle={Information Processing In Medical Imaging},
  year={2023}
}