Deep learning framework for retinal vessel segmentation from Doppler holography images.
This repository contains the training and evaluation code used in the article "Improving segmentation of retinal arteries and veins using cardiac signal in doppler holograms" accepted at ISBI2026 (IEEE, Arxiv). It provides implementations of several segmentation architectures, loss functions, metrics, and utilities for reproducible experiments on holographic Doppler imaging data.
The corresponding dataset is available on Hugging Face: https://huggingface.co/datasets/DigitalHolography/
.
├── models/ # Segmentation model implementations
├── losses.py # Loss functions
├── metrics.py # Evaluation metrics
├── model_utils.py # Utils for models loading and saving
├── data_utils.py # Utils for dataset loading and data visualization
├── requirements.txt # Python dependencies
└── model_benchmark.ipynb # Training and evaluation notebook
pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu126
pip install -r requirements.txtThe main entry point is the model_benchmark.ipynb notebook, which provides examples for:
- Loading the dataset
- Training segmentation models
- Evaluating performance metrics
- Comparing different input modalities
The repository is designed to work with the Retinal Vessel Segmentation from Holographic Doppler Imaging dataset, which contains:
- Power Doppler images (M0)
- Artery/vein segmentation masks
- Correlation maps
- Diasys images
These modalities allow the study of how temporal information derived from the cardiac cycle can improve retinal artery-vein segmentation.
The dataset loading is fully handled in the notebook.
If you use this repository or the dataset in your work, please cite:
M. Dubosc et al., "Improving Segmentation of Retinal Arteries and Veins Using Cardiac Signal in Doppler Holograms," 2026 IEEE 23rd International Symposium on Biomedical Imaging (ISBI), London, United Kingdom, 2026, pp. 1-5, doi: 10.1109/ISBI61048.2026.11515426.