REFUGE2 — Retinal Fundus Glaucoma Challenge 2

2000 fundus images for glaucoma classification, optic disc/cup segmentation, and fovea localisation. Extends REFUGE1.

At a glance

Field	Value
Short name	`refuge2`
Full name	REFUGE2 — Retinal Fundus Glaucoma Challenge 2
Publication date	2020-07-20
Date basis	Repository deposit
Publication date precision	day
Publication date evidence	refuge.grand-challenge.org/Home2020
Publication date source field	Official REFUGE2 Updates: Validation dataset is released
Publication date reviewed	2026-09-11
Date notes	Date applies to the validation dataset announced on the official REFUGE2 challenge page.
Primary category	`fundus`
Resource role	`current_dataset`
Dataset family	`refuge2`
Contained modalities	fundus
Tasks	classification, segmentation
Primary reported quantity	2,000 images
Classes	2 (non-glaucoma, glaucoma)
Splits	train, val, test
Size	1.2 GB
Source-stated terms	Research use (Grand Challenge)
Normalized terms	`research-only`
Descriptive screening label	Research or challenge restriction recorded; check source
Terms scope	`challenge_participation`
Access friction	`anonymous_direct`
Route backend	Google Drive
Availability	`available` (checked 2026-07-21)
Acquisition support	`loader_implemented_not_live_tested`
Legacy sample-loader status	Standard loader included

Reported quantities

Role	Count	Unit	Scope	Basis	Evidence
Primary	2,000	`images`	Primary quantity reported in the reviewed catalog source	`legacy_catalog_field`	drive.google.com/file

Counts retain their source-reported units. Additional rows can describe components, paired items, or derivative copies and are not automatically added to the primary quantity.

Notes

Current release containing all 1,200 REFUGE 2018 images plus 800 additional images. REFUGE 2018 is retained as version history below and is not counted separately.

Documented relationships

These links record source-supported lineage or overlap, not merely similar modality tags.

fundus_domain_generalization is derived from this record: The official description states that the benchmark is based on the original REFUGE images, now represented by the current REFUGE2 record, plus Drishti-GS, ORIGA, and RIGA. (evidence)
lmod_plus is derived from this record: The LMOD+ project page lists nine component datasets, including the five cataloged targets represented by these edges. (evidence)
mm_retinal_reason is derived from this record: The version-pinned official dataset card lists this record among the CFP or OCT sources used to construct MM-Retinal-Reason. (evidence)
refuge1_multirater is derived from this record: The source provides seven-rater annotations for the 1,200 original REFUGE images incorporated into the current REFUGE2 release. (evidence)
smdg is derived from this record: The official SMDG source table lists this catalog record among the 19 standardized source domains. (evidence)

Access information and download

CLI
Python

# Read-only preflight
eyehub download refuge2 --data-dir ./data --dry-run --json

# Download, only when preflight reports supported behavior
eyehub download refuge2 --data-dir ./data

from eyedatahub.acquisition import preflight_dataset
from eyedatahub.datasets.registry import REGISTRY

ds = REGISTRY.get_dataset('refuge2')
print(preflight_dataset(ds, './data'))  # no download

Upstream page: drive.google.com/file

Loader example

This entry includes a standard DatasetSample loader.

from pathlib import Path
from eyedatahub.datasets.registry import REGISTRY

data_dir = Path('~/.eyedatahub/data').expanduser()
ds = REGISTRY.get_dataset('refuge2')
samples = ds.load(data_dir, split='test')
for s in samples[:5]:
    print(s.sample_id, s.label, s.image_path)

Citation

BibTeX
Plain text

@misc{refuge2,
  title  = { REFUGE2 — Retinal Fundus Glaucoma Challenge 2 },
  note   = { Fang H. et al., 'REFUGE2 Challenge: Treasure for Multi-Domain Learning in Glaucoma Assessment', MIA 2022 },
  year   = { 2022 },
  url    = { https://drive.google.com/file/d/1DspRzDqypeBOxZnWPQxmXprNVmJwkBRJ/view },
}

Fang H. et al., 'REFUGE2 Challenge: Treasure for Multi-Domain Learning in Glaucoma Assessment', MIA 2022.

Source-stated terms

Raw source string: Research use (Grand Challenge)
Normalized category: research-only
Apparent scope: challenge_participation
Descriptive screening label: Research or challenge restriction recorded; check source

⚠️ Source-stated terms, scope, and normalized labels are curation metadata, not legal advice or a permission finding. Review the current official source before transfer or reuse.

Similar resources by shared modality

airogs: AIROGS: AI for Robust Glaucoma Screening (113,893 images, cc-by-nc-nd)
multieye: MultiEYE: OCT-Enhanced Fundus Multi-Disease Benchmark (103,959 images, mit)
eyecare_100k: Eyecare-100K: Multimodal Ophthalmology VQA Corpus (102,000 question answer pairs, unknown)
justraigs: JustRAIGS: Just Referral AI Glaucoma Screening Dataset (101,442 images, cc-by-nc-nd)
eyepacs: EyePACS — Diabetic Retinopathy Detection (Kaggle 2015) (88,702 images, research-only)
angioreport: AngioReport Fundus Angiography Report Dataset (55,361 images, unknown)
ffa_ir: FFA-IR Medical Report Dataset (47,247 images, unknown)
mfiddr: MFIDDR: Multi-Field Imaging Dataset for Diabetic Retinopathy (34,452 images, mit)

At a glance​

Reported quantities​

Notes​

Documented relationships​

Access information and download​

Other documented locations​

Loader example​

Citation​

Source-stated terms​

Similar resources by shared modality​