VAN-GAN provides an accessible and efficient deep learning framework for the segmentation of vascular networks in 3D images — without requiring annotated ground truth labels.
As 3D biomedical imaging improves in resolution and accessibility, segmenting vascular networks remains a major bottleneck due to manual annotation requirements. VAN-GAN addresses this using a fully unsupervised deep learning approach based on image-to-image translation.
It adapts and extends the CycleGAN framework to translate between real photoacoustic images and synthetic vessel labels using domain-consistent constraints and 3D residual networks.
- 3D Deep Residual U-Net for realistic vascular structure segmentation
- CycleGAN-style architecture for unpaired domain translation
- No identity loss for simplified training
- Synthetic training images eliminate dependence on manual labels
- Topological & structural constraints for accurate domain alignment
- Sliding window inference for high-resolution image volumes
Clone the repository:
git clone https://github.com/psweens/VAN-GAN.gitInstall dependencies:
pip install opencv-python scikit-image tqdm tensorflow_addons joblib matplotlibEnsure you have:
- Python 3.9+
- TensorFlow 2.10.1
- CUDA 11.2.2 and cuDNN 8.1.0.77
We recommend using a conda environment for clean setup. Follow TensorFlow GPU install guide for compatibility.
| Tool | Version |
|---|---|
| Ubuntu | 22.04.2 LTS |
| Python | 3.9.16 |
| TensorFlow | 2.10.1 |
| Cuda Toolkit | 11.2.2 |
| cuDNN | 8.1.0.77 |
| OpenCV | 4.7.0.72 |
| scikit-image | 0.20.0 |
| tqdm | 4.65.0 |
| tf-addons | 0.20.0 |
| joblib | 1.2.0 |
| matplotlib | 3.7.1 |
A paired dataset of simulated photoacoustic images and synthetic vascular segmentations is available:
📦 Download via University of Cambridge Repository
This dataset is ideal for training VAN-GAN in an unsupervised manner and validating predictions against known vascular structures.
Developed by Paul W. Sweeney.
Community contributions are welcome! If you're using VAN-GAN in your work, feedback and pull requests are encouraged.
If you use VAN-GAN in your research, please cite:
Unsupervised Segmentation of 3D Microvascular Photoacoustic Images Using Deep Generative Learning
Paul W. Sweeney et al., Advanced Science, 2024.
Licensed under the MIT License.
Original CycleGAN implementation adapted from A.K. Nain's TensorFlow example.