FlowVM-Net

This is the official code repository for "FlowVM-Net: Enhanced Vessel Segmentation in X-Ray Coronary Angiography Using Temporal Information Fusion", which is accpeted by Journal of Imaging Informatics in Medicine as a original paper!

Framework Overview

Figure 1: The overall architecture of FlowVM-Net combining spatial features with temporal information

0. Environments.

conda create -n flowvmnet python=3.8
conda activate flowvmnet
pip install torch==1.13.0 torchvision==0.14.0 torchaudio==0.13.0 
pip install packaging
pip install timm==0.4.12
pip install pytest chardet yacs termcolor
pip install submitit tensorboardX
pip install triton==2.0.0
pip install causal_conv1d==1.0.0  # causal_conv1d-1.0.0+cu118torch1.13cxx11abiFALSE-cp38-cp38-linux_x86_64.whl
pip install mamba_ssm==1.0.1  # mmamba_ssm-1.0.1+cu118torch1.13cxx11abiFALSE-cp38-cp38-linux_x86_64.whl
pip install scikit-learn matplotlib thop h5py SimpleITK scikit-image medpy yacs opencv-python

The .whl files of mamba_ssm could be found here. The .whl files of causal_conv1d could be found here.

1. Prepare the dataset.

Data Format

├── './data/XCAV_DIAS/'
    ├── training
        ├── images
            ├── image_s000002_i0.png
            ├── image_s000002_i1.png
            └── ... # image_{sequence_id}_i{frame_index}.png
        ├── labels
            ├── image_s000002_i1.png
            └── ... # same filename as target frame is preferred
    ├── validation
        ├── images
        ├── labels
    ├── test
        ├── images
        ├── labels

The loader groups frames by the part before _i{frame_index}. For example, image_s000002_i1.png through image_s000002_i5.png are treated as one temporal sequence. If the label is sequence-level, such as image_s000002.png, it is matched only with the window ending at the final frame. With num_frames=2, that sample is i4 + i5 -> image_s000002.png, not i0 + i1 or i1 + i2. Frame-level labels such as image_s000002_i5.png are still supported. The old train/val/test and masks naming scheme is still supported for compatibility.

2. Train the FlowVM-Net.

• The weights of the pre-trained VMamba could be downloaded here.
• The weights of the pre-trained Optical Flow model could be downloaded here. After that, the pre-trained weights should be stored in './pre_trained_weights/'.

bash train.sh
# or
python train.py --data_path /path/to/XCAV_DIAS --gpu_id 0

• After trianing, you could obtain the outputs in ./results/

3. Test the FlowVM-Net.

python testing.py --data_path /path/to/XCAV_DIAS --checkpoint_path /path/to/best.pth --gpu_id 0

4. Citation

Please cite the paper as follows if you use the code from FlowVM-Net:

@article{wei2025flowvm,
  title={FlowVM-Net: Enhanced Vessel Segmentation in X-Ray Coronary Angiography Using Temporal Information Fusion},
  author={Guangyu Wei and Xueying Zeng and Qing Zhang},
  journal={Journal of Imaging Informatics in Medicine},
  year={2025},
  publisher={Springer},
  doi={10.1007/s10278-025-01732-y}
}