LUNet: deep learning for the segmentation of arterioles and venules in high resolution fundus images

Jonathan Fhima • Jan Van Eijgen • Marie-Isaline Billen Moulin-Romsée • Heloïse Brackenier• Hana Kulenovic

Valérie Debeuf Marie Vangilbergen• Moti Freiman• Ingeborg Stalmans• Joachim A Behar

Installation

First, clone this repository and run install the environment:

cd LUNet
python -m venv lunet_env
source lunet_env/bin/activate
pip install -r requirements.txt

Data preparation

Download the UZLF (aka Leuven-Haifa) dataset from the official website (https://rdr.kuleuven.be/dataset.xhtml?persistentId=doi:10.48804/Z7SHGO), and organize it as follows:

LUNet
├── Databases
│   ├── UZLF_TRAIN
│   │   ├── images
│   │   ├── artery
│   │   ├── veins
│   ├── UZLF_VAL
│   │   ├── images
│   │   ├── artery
│   │   ├── veins
│   ├── UZLF_TEST
│   │   ├── images
│   │   ├── artery
│   │   ├── veins

After install the external datasets from PVBM (https://pvbm.readthedocs.io/) using the following commands:

cd LUNet
source lunet_env/bin/activate
python -u install_pvbm_datasets.py

Training

Run the following command: (Original LUNet model have been trained using 8 A100-40gb GPUs and have been trained with a slightly different version of the UZLF dataset which can explain minor differences in performance).

cd LUNet
source lunet_env/bin/activate
python -u main.py Databases/ lunet_model

Evaluation

Run with test-time data augmentation (higher performance):

cd LUNet
source lunet_env/bin/activate
python -u eval_all.py Databases/ lunet_model --use_TTDA --datasets_test UZLF_VAL UZLF_TEST CropHRF INSPIRE
#Remove --use_TTDA if you dont want to use test time data augmentation during inference
#You can change the list of the test dataset

Run without test-time data augmentation (faster inference):

cd LUNet
source lunet_env/bin/activate
python -u eval_all.py Databases/ lunet_model --datasets_test UZLF_VAL UZLF_TEST CropHRF INSPIRE

Main Results

Method	UZLF test		LES-AV		INSPIRE-AVR		Cropped HRF
	Artery	Vein	Artery	Vein	Artery	Vein	Artery	Vein
LUNet	82.0	84.5	82.3	84.8	73.6	77.5	78.1	80.4

LICENSE

This project is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License, see LICENSE file, which prohibits commercial use.

Citation

If you find this code or data to be useful for your research, please consider citing the following papers.

@article{fhima2024lunet,
    title={LUNet: deep learning for the segmentation of arterioles and venules in high resolution fundus images},
    author={Fhima, Jonathan and Van Eijgen, Jan and Moulin-Roms{\'e}e, Marie-Isaline Billen and Brackenier, Helo{\"\i}se and Kulenovic, Hana and Debeuf, Val{\'e}rie and Vangilbergen, Marie and Freiman, Moti and Stalmans, Ingeborg and Behar, Joachim A},
    journal={Physiological Measurement},
    volume={45},
    number={5},
    pages={055002},
    year={2024},
    publisher={IOP Publishing}
}

@article{van2024leuven,
    title={Leuven-Haifa High-Resolution Fundus Image Dataset for Retinal Blood Vessel Segmentation and Glaucoma Diagnosis},
    author={Van Eijgen, Jan and Fhima, Jonathan and Billen Moulin-Roms{\'e}e, Marie-Isaline and Behar, Joachim A and Christinaki, Eirini and Stalmans, Ingeborg},
    journal={Scientific Data},
    volume={11},
    number={1},
    pages={257},
    year={2024},
    publisher={Nature Publishing Group UK London}}

@InProceedings{10.1007/978-3-031-25066-8_15,
    author="Fhima, Jonathan and Eijgen, Jan Van and Stalmans, Ingeborg and Men, Yevgeniy and Freiman, Moti and Behar, Joachim A.",
    title="PVBM: A Python Vasculature Biomarker Toolbox Based on Retinal Blood Vessel Segmentation",
    booktitle="Computer Vision -- ECCV 2022 Workshops",
    year="2023",
    publisher="Springer Nature Switzerland",
    address="Cham",
    pages="296--312",
    isbn="978-3-031-25066-8"
    }