HAMIL: High-resolution Activation Maps and Interleaved Learning for Weakly Supervised Segmentation of Histopathological Images

This repository provides the code for "HAMIL: High-resolution Activation Maps and Interleaved Learning for Weakly Supervised Segmentation of Histopathological Images" accepted by TMI 2023.

Usage

1. To obtain the background masks. (If the background is not white regions, skip this step)

train_set_root: training set root, gamma_path: path to save gamma transform for training set, gamma_crf_path: path to save extracted backgrounds for training set,

generate_bg_masks.py --train_set_root train_set_root --gamma_path gamma_path --gamma_crf_path gamma_crf_path

2. Train classification network

train_cls.py --dataset_root dataset_root --gpu 0

3. Train segmentation network

train_seg.py --dataset_root dataset_root --gpu 0

Citation

@article{zhong2023hamil,
  title={HAMIL: High-resolution Activation Maps and Interleaved Learning for Weakly Supervised Segmentation of Histopathological Images},
  author={Zhong, Lanfeng and Wang, Guotai and Liao, Xin and Zhang, Shaoting},
  journal={IEEE Transactions on Medical Imaging},
  year={2023},
  publisher={IEEE}
}

Acknowledgement

The code of DeepLabv3+ is borrowed from PuzzleCAM