Official implementation of CVI

[NOTE!!]The code will be gradually and continuously opened!

In this work we proposed a novel Conditional Virtual Imaging (CVI) framework based on pre-trained diffusion model and our dual-consistency learning strategy which can generate authentic vessel images with few training data and achieve good performance in few-shot vessel segmentation tasks. Our model is implemented on vessel segmentation tasks of different datasets including coronary artery and retinal vessel under one-shot and five-shot situations. The promising results compared with other methods show our great superiority.

pip install -r requirements.txt

python experiments/train_cvi.py --config configs/coronary_1shot.yaml

python experiments/evaluate.py --config configs/coronary_1shot.yaml --checkpoint path/to/checkpoint.pth

This repository provides the official implementation of CVI in the following papers:

Conditional Virtual Imaging for Few-Shot Vascular Image Segmentation
Yanglong He, Rongjun Ge, Hui Tang, Yuxin Liu, Mengqing Su, Jean-Louis Coatrieux, Huazhong Shu, Yang Chen*, Yuting He*
Southeast University
IEEE Transactions on Medical Imaging

If you use this code or use our pre-trained weights for your research, please cite our papers:

@article{he2025conditional,
  title={Conditional Virtual Imaging for Few-Shot Vascular Image Segmentation},
  author={He, Yanglong and Ge, Rongjun and Tang, Hui and Liu, Yuxin and Su, Mengqing and Coatrieux, Jean-Louis and Shu, Huazhong and Chen, Yang and He, Yuting},
  journal={IEEE Transactions on Medical Imaging},
  year={2025},
  publisher={IEEE}
}