🔬 PDSeg: Patch-Wise Distillation and Controllable Image Generation for Weakly-Supervised Histopathology Tissue Segmentation

Welcome! 👋 This repository contains the official implementation of our ICASSP 2025 paper:
"PDSeg: Patch-Wise Distillation and Controllable Image Generation for Weakly-Supervised Histopathology Tissue Segmentation".

We propose a novel framework that combines patch-wise knowledge distillation with controllable image generation to push the boundaries of weakly-supervised tissue segmentation in histopathology images.

---

🖼️ Overview

---

📊 Click to view: Comparison on LUAD-HistoSeg and BCSS-WSSS

---

⚙️ Installation and Requirements

We are tested under:

python 3.11

If you want to install a custom environment for this code, you can run the following using conda:

conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
conda install matplotlib

pip install timm==0.6.13
pip install opencv-python
pip install einops
pip install scikit-learn
pip install scikit-image

📁 Datasets and Our checkpoints

You can get the LUAD-HistoSeg and BCSS-WSSS dataset from here and our checkpoints.

If your dataset is in a different folder, make a soft-link from the target dataset to the data folder. We expect the following tree:

data/
    BCSS-WSSS/
      training/
      test/
      val/
    LUAD-HistoSeg/
checkpoint/
    bcss_baseline/
      best_model.pth
    luad_baseline/
      best_model.pth
... other files 

🏋️ Training and Evaluation

for LUAD-HistoSeg

bash run_luad.sh

for BCSS-WSSS

bash run_bcss.sh

📝 Citation

If you find this work useful in your research, please cite our paper:

@inproceedings{li2025pdseg,
  title={PDSeg: Patch-Wise Distillation and Controllable Image Generation for Weakly-Supervised Histopathology Tissue Segmentation},
  author={Li, Wei-Hua and Hsieh, Yu-Hsing and Yang, Huei-Fang and Chen, Chu-Song},
  booktitle={ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={1--5},
  year={2025},
  organization={IEEE}
}

📄 License

This project is released under a custom license.
Please see the LICENSE file for the full terms and conditions.

For academic or commercial use, please contact the authors.

---

Stay tuned for code release and updates!
📬 Questions or feedback? Feel free to open an issue or contact us.