ProCNS: Progressive Prototype Calibration and Noise Suppression for Weakly-Supervised Medical Image Segmentation

The official implementation of the paper: ProCNS: Progressive Prototype Calibration and Noise Suppression for Weakly-Supervised Medical Image Segmentation

🔔 News

• 2024-12-23, 🎉🎉 Our paper "ProCNS: Progressive Prototype Calibration and Noise Suppression for Weakly-Supervised Medical Image Segmentation" has been accepted by IEEE Journal of Biomedical and Health Informatics (JBHI).

💡 Framework

🌟 Contributions (Why Cite This Work?)

This work provides insights for researchers exploring:

1. Boundary Matters in Medical Weakly-Supervised Medical Image Segmentation
ProCNS systematically relieves error accumulation in ambiguous boundary regions. Our progressive prototype calibration improves edge segmentation by 3.25-5.43% DSC through:

• PRSA Loss: Joint spatial-semantic affinity modeling
• ANPM Module: Dynamic noise masking

2. Foundation Model Meets Weak Supervision
Pioneering integration of SAM-Med2D into WSS:

• Prove that leveraging foundation models to assist WSS is a promising direction
• Generate high-quality pseudo-labels via point prompts (sparse labels) (Fig.7)
• Achieve 93.30 DSC when combined with ProCNS (Table 9)

3. Unified Medical WSS Benchmark Released
The Unified WSS benchmark covering:

✅ 6 imaging modalities
✅ 3 annotation types
✅ Standard evaluation protocols

Datasets

Download the Processed Datasets

Modality	Fundus	OCTA	Endoscopy	Cardiac MRI	Brain Tumor MRI	H&E
Sparse Annotation Format	scribble	point	block	scribble	block	point
Source	Download	Download	Download	Download	Download	Paper
Processed (.h5)	Download	Download	Download	Download	N/A	N/A

N/A will be included in subsequent updates. If you need the processed datasets (.h5) of the BraTS2019 and H&E, please feel free to contact us.

Requirements

Some important required packages are listed below:

• Pytorch 1.10.2
• cudatoolkit 11.3.1
• efficientnet-pytorch 0.7.1
• tensorboardx 2.5.1
• medpy 0.4.0
• scikit-image 0.19.3
• simpleitk 2.1.1.2
• Python >= 3.9

Usage

1. Clone this project

git clone https://github.com/LyxDLiI/ProCNS.git
cd ProCNS/code

2. Create a conda environment

conda env create -n procns -f procns.yaml
conda activate procns
pip install tree_filter-0.1-cp39-cp39-linux_x86_64.whl

3. Pre-processing

Data preprocessing includes normalizing all image intensities to between 0 and 1, while data augmentation includes randomly flipping images horizontally and vertically as well as rotation (spanning from -45° to 45°).

4. Train the model

run.sh

5. Test the model

test.sh

6. Result

7. Visualization

Acknowledgement

• WSL4MIS
• FedICRA
• FedLPPA

Citation

If you find ProCNS useful in your research, please consider citing:

@article{liu2024procns,
  title={ProCNS: Progressive Prototype Calibration and Noise Suppression for Weakly-Supervised Medical Image Segmentation},
  author={Liu, Yixiang and Lin, Li and Wong, Kenneth KY and Tang, Xiaoying},
  journal={IEEE Journal of Biomedical and Health Informatics},
  year={2024},
  publisher={IEEE}
}

If you have any questions, please feel free to contact us.