REYIA bundles ENRICH, a dataset of retinal fundus images collected under KU Leuven Hospital Helsinki Committee approval (S66963). Access requests are logged to keep a record of who is using the data and for what purpose, per that approval's terms.

By requesting access you agree to use ENRICH (images and annotations) for non-commercial purposes only, per its KU Leuven data-sharing protocol. REYIA's own annotations and code are Apache 2.0, and source images from the other subsets follow their own original licenses (see below). You agree to cite the REYIA paper and each original subset's own citation in any resulting publication.

REYIA

Jonathan Fhima • Jan Van Eijgen • Lennert Beeckmans • Thomas Jacob • Moti Freiman
Luis Filipe Nakayama • Ingeborg Stalmans • Chaim Baskin • Joachim A. Behar

REYIA is a curated set of 589 retinal fundus images annotated with AV (artery/vein) blood vessel segmentations using the open-access Lirot.ai software. To enhance diversity, REYIA includes manually segmented images from nine sources: FIVES, TREND, GRAPE, MESSIDOR, MAGRABIA, PAPILA, mBRSET, AV-WIDE, and ENRICH, a new dataset of 111 images collected for this study. AV-WIDE, which originally shipped only skeletonized vessel traces, was reannotated here with complete AV segmentations.

This repository ships the annotations for all 589 images, plus the source images themselves wherever their license allows redistribution. For the rest, it ships a script that reconstructs the exact image REYIA annotated from your own raw download.

License

REYIA's own annotations and the code in scripts/ are licensed under Apache 2.0 (see LICENSE), with one exception: ENRICH — both the images and their annotations — is non-commercial use only, per the KU Leuven Hospital data-sharing protocol under which it was collected (Helsinki approval S66963). Source images from the other eight datasets keep their own original licenses:

Subset	Bundled here	License
FIVES	Yes	CC BY 4.0
GRAPE	Yes	CC0 1.0
TREND	Yes	CC BY 4.0
ENRICH	Yes	Non-commercial use only — KU Leuven data-sharing protocol (own data)
PAPILA	No	figshare license field reads GPL 3.0+ — likely a depositor mis-click, but not corrected as of this writing
MAGRABIA	No	Deep Blue Data tag is CC BY-NC 4.0; scope relative to the rest of RIGA unresolved
MESSIDOR	No	Explicit no-redistribution terms
mBRSET	No	PhysioNet Data Use Agreement, credentialed access only
AV-WIDE	No	Explicit no-redistribution terms

Annotations for every subset other than ENRICH are released here under Apache 2.0 regardless of the source image's own license. ENRICH's annotations follow ENRICH's own non-commercial restriction above.

Download & build

python3 -m venv reyia_env && source reyia_env/bin/activate
pip install huggingface_hub
huggingface-cli download JonathanFhima/REYIA --repo-type dataset --local-dir REYIA
cd REYIA

FIVES, GRAPE, TREND, and ENRICH are already bundled — use images/<subset>/ and annotations/<subset>/ directly.

For the other five, download each one and extract it — as is, no need to pick files out — into raw_downloads/<subset>/ (you don't need all five; subfolders inside are fine, everything is found recursively):

Subset	Download	Extract into	What's in there
PAPILA	https://figshare.com/articles/dataset/PAPILA/14798004	`raw_downloads/papila/`	the `.jpg` fundus photos
MAGRABIA	https://deepblue.lib.umich.edu/data/concern/data_sets/3b591905z	`raw_downloads/magrabia/`	keep the `MagrabiFemale/` and `MagrabiaMale/` folders as-is — their names are how the script tells images apart
MESSIDOR	https://www.adcis.net/en/third-party/messidor2/ (registration required)	`raw_downloads/messidor/`	just the `Base11` package (one of 9 zips) — the `.tif` files
mBRSET	https://physionet.org/content/mbrset/1.0/ (credentialed access)	`raw_downloads/mbrset/`	the `images/` folder from the download
AV-WIDE	https://people.duke.edu/~sf59/Estrada_TMI_2015_dataset.htm	`raw_downloads/av-wide/`	the `AV-WIDE/` folder of `.mat` files

pip install -r scripts/build/requirements.txt
bash scripts/build/build_reyia.sh

This writes each reconstructed image into images/<subset>/, matched to its reyia_id. Output is visually identical to REYIA's own copy, though not always byte-identical (resampling rounding — expected, not an error).

Verified pixel-exact for PAPILA, MAGRABIA, MESSIDOR, and AV-WIDE. mBRSET: 25/30 exact; the other 5 don't resolve under any crop/threshold tried — their PhysioNet source files most likely changed since REYIA's download.

Ethical Consent

ENRICH. Approved by the Helsinki Committee at KU Leuven Hospital (approval S66963); informed consent waived, all identifiable patient information removed.

GRAPE. Registered clinical study (ClinicalTrials.gov NCT04718532), approved by Ethics Committee of ZJU-2 (No Y2020-1027). https://www.nature.com/articles/s41597-023-02424-4

MESSIDOR. Free for research/educational use; redistribution and unauthorized commercial use prohibited; publications must acknowledge LaTIM laboratory and Messidor program partners. https://www.adcis.net/en/third-party/messidor2/

PAPILA. Collected at Hospital General Universitario Reina Sofía (Murcia, Spain), 2018–2020, under hospital Ethics Committee approval with patient informed consent. https://www.nature.com/articles/s41597-022-01388-1

MAGRABIA. Ethical approval obtained for the de-identified dataset from University of Waterloo, Canada. https://www.spiedigitallibrary.org/conference-proceedings-of-spie/10579/2293584/Retinal-fundus-images-for-glaucoma-analysis-the-RIGA-dataset/10.1117/12.2293584.full

FIVES. Approved by the Medical Ethics Committee of the Second Affiliated Hospital of Zhejiang University; written informed consent from every participant; registered on ClinicalTrials.gov (NCT04718532). https://www.nature.com/articles/s41597-022-01564-3

AV-WIDE. Acquired at Duke University Medical Center under Duke IRB protocols Pro00015512 and Pro00056311. https://people.duke.edu/~sf59/estradaTMI15.pdf

TREND. Approved by the Ethical Committee of the Faculty of Medicine, University of Montenegro (Protocol No. 2487/4 and 3824/4 / 03/01-11417/1). https://pmc.ncbi.nlm.nih.gov/articles/PMC8301647/

mBRSET. Approved by the Institutional Review Board of IPTAN (protocol CAAE 64219922.3.0000.9667); all identifiable patient information removed; written consent obtained for image capture and open publication. https://www.nature.com/articles/s41597-025-04627-3

Benchmark reference

If you train a model on REYIA (or on data augmented with it), these are our own reference numbers — artery/vein Dice score, averaged — across the same local, external, and out-of-distribution (OOD) test sets, so you can compare directly instead of re-running every baseline yourself. ∆ rows show the improvement RLAD adds on top of each backbone. † marks results with data leakage during training.

scripts/eval/dice_score.py is the exact scorer used to produce these numbers — artery is the mask's red channel, vein is the blue channel, both nonzero after thresholding, so magenta crossing pixels count toward both.

pip install -r scripts/eval/requirements.txt
python3 scripts/eval/eval_example.py --mask annotations/<subset>/reyia_XXXX.png

This scores a real REYIA mask against itself (prints 1.0 across the board, as a sanity check).

Citation

If you use REYIA, please cite the paper it was introduced in:

Fhima, J., Van Eijgen, J., Beeckmans, L., Jacob, T., Freiman, M., Nakayama, L.F., Stalmans, I., Baskin, C., Behar, J.A. Enhancing Retinal Vessel Segmentation Generalization via Layout-Aware Generative Modelling. arXiv:2503.01190, 2025. https://arxiv.org/abs/2503.01190

Code (RLAD): https://github.com/aim-lab/RLAD

Please also cite the original dataset(s) any images you use come from, per their own citation requirements (see the Ethical Consent section above).

Downloads last month: 1

Total file size:

1.35 GB

Paper for JonathanFhima/REYIA

Enhancing Retinal Vessel Segmentation Generalization via Layout-Aware Generative Modelling

Paper • 2503.01190 • Published Apr 5, 2025