HRF — High-Resolution Fundus Image Database

Manual retinal blood-vessel segmentation ground truth for 45 high-resolution color fundus photographs (3504×2336) from the Pattern Recognition Lab (CS5) and Dept. of Ophthalmology, FAU Erlangen-Nürnberg, with Brno University of Technology and the Eye Clinic Zlín — Budai et al., Int. J. Biomedical Imaging 2013.

• Modality: color fundus photography (2D RGB), 3504×2336
• Organ: retina / eye
• Target: binary blood-vessel segmentation
• Cases: 45 — 15 healthy · 15 glaucomatous · 15 diabetic retinopathy
• License: CC BY 4.0
• Source: FAU HRF homepage (official, author-hosted)

Scope — please read. This repository contains only the segmentation component of HRF. The 36-image Image Quality Assessment set hosted on the same FAU page (18 good/bad pairs, different resolutions, no segmentation ground truth, separate citation — Köhler et al., CBMS 2013) is not included: it is a different dataset that shares a download page. A naive "download HRF" yields 45+36 = 81 images; this repo is the 45 segmentation images.

Splits

There is no official train/test split. Neither the FAU page nor the Budai paper defines one, so all 45 cases are published as a single train split.

The split most often seen in the literature is a community convention from Orlando et al. 2017 (IEEE TBME 64(1):16–27) — the first 5 of each diagnosis category for training (15) and the remaining 30 for testing. It is not author-defined. Reconstruct it exactly with case_index <= 5 (train) vs case_index >= 6 (test).

Columns

Column	Type	Notes
image_id	string	Case stem, e.g. 01_h, 07_g, 15_dr — the pairing key
image	Image (RGB)	3504×2336 fundus photograph. Original JPEG bytes, stored verbatim (no re-encode)
mask	Image (L)	Ground truth. Manual binary vessel segmentation, {0, 255}, 3504×2336
fov_mask	Image (L)	Field-of-view / camera-aperture mask, {0, 255}. Auxiliary — not a segmentation target
subset	string	healthy | glaucoma | diabetic_retinopathy
case_index	int32	1–15 within the subset (parsed from the filename prefix)
vessel_fraction	float32	Fraction of pixels labelled vessel (0.051–0.106, mean 0.077)
fov_fraction	float32	Fraction of pixels inside the FOV (~0.845)
od_center_x_a, od_center_y_a	int32	Optic-disc ("papilla") centre, Expert A
od_vessel_origin_x_a, od_vessel_origin_y_a	int32	Central-vessel origin, Expert A
od_diameter_a	int32	Optic-disc diameter in px, Expert A
od_center_x_b, od_center_y_b	int32	Optic-disc centre, Expert B
od_vessel_origin_x_b, od_vessel_origin_y_b	int32	Central-vessel origin, Expert B
od_diameter_b	int32	Optic-disc diameter in px, Expert B

For binary vessel segmentation use mask > 0 (masks are already clean two-valued).

The optic-disc columns are coordinates, not masks. FAU ships an optic_disk_centers.xls "Optic Disk Goldstandard" that contains localisation annotations only — there is no optic-disc segmentation in HRF. They are carried here as metadata for localisation/registration use.

Ground truth

mask is the manual vessel segmentation from manual1/. It is the only tier the Budai paper evaluates against (Tables 4–5, per-subset Se/Sp/Acc), and the FAU page describes it as produced by "a group of experts working in the field of retinal image analysis and clinicians from the cooperated ophthalmology clinics" — a single expert-consensus set, so there is no rater to choose.

HRF has no second-observer segmentation. (The Budai paper's remark about "a second manual segmentation made by a human observer" refers to DRIVE and STARE, the external databases it compares against — not to HRF.)

Mask normalisation applied here

42 of the 45 source vessel masks are strictly {0, 255}. Three — 11_h, 12_h, 13_h — carry anti-aliased grey edges (175/215/203 distinct values); they are exactly the three files stored uncompressed rather than PackBits upstream, i.e. a different export path. Affected pixels: 599 / 1584 / 1119 out of 8,185,344 (0.007–0.019%). All masks were binarised at ≥128 so every case ships clean {0, 255}. The ≥128 threshold (rather than >0) keeps those three geometrically consistent with the other 42 instead of gaining a ~1 px anti-alias fringe.

FOV masks are stored 3-channel RGB (R==G==B) at the source; they are reduced to single-channel here.

Provenance, naming and cross-dataset overlap

• Provenance: official, author-hosted FAU archives (all.zip, 76,317,613 B, byte-size verified). Counts match the paper exactly: 15/15/15 = 45. ⚠️ Third-party re-hosts vary in fidelity — one HF mirror carries 90 downscaled rows with no vessel GT. This repo is built from the FAU originals.
• Faithful naming: yes, with the IQA-set exclusion noted in the scope box above.
• Ground-truth tier: single expert-consensus vessel set; no second observer.
• Cross-dataset overlap: NONE. HRF shares no images, archives or lineage with DRIVE, STARE, CHASE_DB1, IDRiD, PAPILA, RAVIR, RITE or Messidor — different acquisition sites, cameras and resolutions. No cross-reference ID exists or is needed. (Note that third-party re-annotations of these same 45 images do exist — e.g. HRF-AV artery/vein labels, HRF-Seg+ — and overlap HRF 100%; do not benchmark those alongside this repo.)

Citation

@article{budai2013robust,
  title   = {Robust Vessel Segmentation in Fundus Images},
  author  = {Budai, Attila and Bock, R{\"u}diger and Maier, Andreas and
             Hornegger, Joachim and Michelson, Georg},
  journal = {International Journal of Biomedical Imaging},
  volume  = {2013},
  pages   = {154860},
  year    = {2013},
  doi     = {10.1155/2013/154860}
}

Companion database paper: Odstrcilik, J. et al. "Retinal vessel segmentation by improved matched filtering: evaluation on a new high-resolution fundus image database." IET Image Processing 7(4):373–383, 2013. doi:10.1049/iet-ipr.2012.0455