Cross-view Generalized Diffusion Model for Sparse-view CT Reconstruction

🚀 Updates

• The code of training & test are released.
• The weights of trained models are released.

⭐ Highlights of CvG-Diff

• CvG-Diff explicit models the angular sparsity artifacts within image-domain as a deterministic degradation process to build a generalied diffusion model. In this way, CvG-Diff harness the simultaneous reconstruction capability across different sparsity regimes, enabling high-quality iterative reconstruction with few sampling steps.
• To better suit the sparsity artifacts in CT images, CvG-Diff integrates: 1) an Error-Propagating Compositve Training (EPCT) strategy to suppress the artifact propagation across different sparsity regimes; 2) a Semantic-Proritized Dual-Phase Sampling (SPDPS) strategy that emphasize overall anatomical structural correctness before detail refinement during iterative sampling.
• CvG-Diff achieves SoTA performance in sparse-view CT reconstruction across various sparsity levels, enabling few-step high-quality reconstruction.

🔨 Environment

• Download the code, create a conda environment, and install the required packages in requirements.txt by running the following commands:

  git clone https://github.com/xmed-lab/CvG-Diff.git
  cd CvG-Diff
  conda create -n CvG-Diff python=3.7
  conda activate CvG-Diff
  pip install -r ./requirements.txt

• torch-radon is required for simulating DRRs and geometry utils, install torch-radon by

1. Download torch-radon from torch-radon

   git clone https://github.com/matteo-ronchetti/torch-radon.git

2. Due to some out-dated Pytorch function in torch-radon, you need to modify code by running

   cd torch-radon
   patch -p1 < path/to/CvG-Diff/torch-radon_fix/torch-radon_fix.patch

3. Install torch-radon by running

   python setup.py install

💻 Prepare Dataset

• Please download original AAPM16 dataset from AAPM and process them with code in ./datasets/preprocess_aapm.py.

🔑 Train and Evaluate

• We provide the checkpoint of trained CvG-Diff the following links, you can download them for direct inference: BaiduNetdisk; Huggingface.
• Run training: Open the train.sh file and config your dataset path and result directory properly, then run

  train.sh

• Run evaluation: Open the test.sh file and config your dataset path and result directory properly, then run

  test.sh

📘 Results

📚 Citation

If you find our paper helps you, please kindly cite our paper in your publications.

@inproceedings{chen2025cvgdiff,
  title={Cross-view Generalized Diffusion Model for Sparse-view CT Reconstruction},
  author={Chen, Jixiang and Lin, Yiqun and Qin, YI and Wang Hualiang and Li, Xiaomeng},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year={2025},
  organization={Springer}
}

🍻 Acknowledge

Our code is built upon Freeseed and ColdDiffusion, many thanks to their authors for sharing their code.