Datasets:

le-dang1haha
/

FundusShift

Purpose. Measure how models for colour fundus screening hold up, in accuracy and calibration, on whole collections never seen in development.
Creators and funding. To be filled in by the authors.

Composition

Instances. One colour fundus photograph per row. 35,280 released of 40,414 inventoried, from 15 collections. Each exclusion and its reason is in metadata/splits_final.csv.
Labels. N (apparently normal), D (diabetic-retinopathy-like), G (glaucoma), plus an open-set split outside these classes. Labels describe the photograph as annotated by the source. They are not confirmed diagnoses.
Splits. train 23,805, validation 2,983, calibration 2,968, test 3,320 (7 held-out collections), test_ood 1,343 (Chákṣu), test_openset 861.
Sample or complete. A curated sample of openly licensed collections. It does not represent any population.
Demographics. Age and sex appear only in PAPILA among the files we hold. Fairness by subgroup cannot be assessed across the benchmark.
Sensitive data. Medical images released by their sources. Some originals carry burned-in text (for example dates or eye tags), visible in samples. Whether any is identifying has not been audited.
Known problems. See Limitations.

Collection. Taken from public releases. Devices, countries and licences are per collection in the card. Consent and ethics approval rest with each source and were not reviewed one by one; this review is still to do.
Preprocessing. Crop to the illuminated field, longest side at most 1024 px, JPEG q95. Original SHA-256 kept.
Duplicates. Exact hash, then vessel-pattern NCC at threshold 0.70, joined with patient IDs into leakage groups that never span two splits.

Intended. Research on domain shift, calibration, selective prediction and open-set referral.
Not intended. Clinical use or any deployment decision.
Risks. Pooled accuracy hides failure by collection and class. Report per collection and class.

Grading scales differ across sources. N means "no DR" in DR-only collections.
Patient identifiers exist only for some collections.
HYGD holds only G and PALM only N in test, so per-source scores there are one-class recall. HRF has 15 images per class.
Within several collections, trivial image statistics predict the class (for example FIVES). Quality matching reduced this only partly and failed on small collections.
Labels are kept as published by each collection; no label was changed. No independent clinical re-grading was done, so the rate of label error is unknown. A pilot with AI raters was too weak on glaucoma to estimate it (kappa with the published labels 0.30 to 0.38).
Pretrained models may have seen held-out collections: FLAIR's list includes FIVES, HRF, PAPILA, Paraguay and Chákṣu.
Confident learning flags are candidates only. The detector found about 57 % of synthetic flips at about 26 % precision.

Hugging Face, private at present: le-dang1haha/FundusShift. No DOI yet.
No single licence. Each image keeps its source licence. OIA-DDR (CC BY-NC-SA 4.0, 12,520 images), DeepDRiD (CC BY-SA 4.0, 2,000) and PAPILA (GPL-3.0-or-later, 420) carry the strongest terms. Published terms allow research use and non-commercial redistribution with attribution, stated changes and the same licence. Use of the full benchmark is non-commercial.
Not legal advice. PAPILA's licence was confirmed from a University of Murcia record because figshare could not be opened; check the figshare field before public release. Commercial use of models trained on OIA-DDR is an open question for its authors.