Dataset Exploration Guide

This guide is for humans and software clients looking for ophthalmology datasets to add to EyeDataHub. It is organized as a practical search map rather than a catalog snapshot. Use it to broaden candidate discovery, structure notes, and decide whether a resource is distinct and sufficiently evidenced for catalog review.

Before adding metadata, check the current catalog:

eyehub search --json
eyehub show <name> --json

Do not infer missing licenses, sample counts, demographics, or access terms. When source material is ambiguous, preserve the ambiguity in notes and mark the license family as unknown until verified.

Exploration Workflow

1. Define the anatomical scope, modality, and task.
2. Search with several synonym groups, including disease names, device names, annotation types, and repository platforms.
3. Record the primary source URL, data paper, challenge page, access platform, license text, and citation before interpreting the dataset.
4. Compare candidate names, sources, publications, sample counts, and images against the existing catalog.
5. Decide whether the candidate is a new dataset, a subset, a mirror, a benchmark split, a derived annotation set, or a duplicate.
6. Add only source-supported metadata.

Anatomical Search Map

Use anatomical terms to avoid overfitting searches to a single disease or modality.

Anatomical focus	Common data targets	Useful search terms
Retina, posterior pole	Color fundus images, ultra-widefield images, OCT, OCTA, vessel masks, lesion labels	retina dataset, retinal image dataset, fundus dataset, posterior pole, diabetic retinopathy images, retinal vessel segmentation
Macula	OCT B-scans, OCT volumes, fundus photographs, OCTA maps, edema labels, AMD grades	macular OCT dataset, age-related macular degeneration dataset, diabetic macular edema, macula segmentation, retinal fluid segmentation
Optic disc and nerve fiber layer	Fundus optic disc crops, cup-disc masks, OCT RNFL scans, visual fields	glaucoma dataset, optic disc segmentation, optic cup segmentation, RNFL OCT dataset, visual field dataset
Choroid	Enhanced-depth OCT, OCTA, ICGA, choroidal thickness maps	choroid OCT dataset, choroidal segmentation, pachychoroid dataset, indocyanine green angiography dataset
Retinal vasculature	Fundus vessel masks, fluorescein angiography, OCTA capillary plexus maps	retinal vessel dataset, vessel segmentation, artery vein classification, OCTA dataset, fluorescein angiography dataset
Anterior chamber and angle	AS-OCT, UBM, slit-lamp gonioscopy, angle closure labels	anterior segment OCT dataset, anterior chamber angle dataset, gonioscopy dataset, ultrasound biomicroscopy eye dataset
Cornea	Slit-lamp images, specular microscopy, confocal microscopy, topography, tomography	corneal dataset, keratoconus dataset, specular microscopy dataset, corneal endothelium dataset, corneal topography dataset
Lens and cataract	Slit-lamp images, retroillumination, fundus quality labels, cataract grades	cataract dataset, lens opacity dataset, slit lamp cataract images, cataract grading dataset
Conjunctiva and sclera	External eye photographs, slit-lamp images, redness and lesion labels	conjunctiva dataset, ocular surface dataset, sclera segmentation, eye redness dataset, pterygium dataset
Eyelid, periocular, orbit	External photographs, periocular biometrics, ptosis labels, thyroid eye disease images	eyelid dataset, periocular dataset, ptosis dataset, orbital disease dataset, thyroid eye disease image dataset
Eye movements and gaze	Eye tracking streams, video, fixation and saccade labels, gaze estimation	eye tracking dataset, gaze estimation dataset, fixation dataset, saccade dataset, ocular motility dataset
Functional vision	Visual fields, electrophysiology, acuity, contrast sensitivity, clinical tabular data	visual field dataset, perimetry dataset, ERG dataset, visual acuity dataset, ophthalmology tabular dataset

Modality And Device Search Map

Search by both modality and device because many dataset pages use vendor or instrument language rather than clinical anatomy.

Modality or device class	Common fields to verify	Useful search terms
Color fundus photography	Field of view, mydriatic status, camera model, laterality, grading scale	color fundus dataset, retinal fundus images, CFP dataset, mydriatic fundus, nonmydriatic fundus
Ultra-widefield fundus imaging	Device, field of view, peripheral lesion labels, montage or raw image status	ultra-widefield retinal dataset, UWF fundus dataset, Optos dataset, peripheral retina dataset
OCT B-scan and volume	Vendor, scan protocol, voxel spacing if available, disease labels, layer or fluid masks	OCT dataset, spectral domain OCT, swept source OCT, retinal OCT volume, OCT segmentation dataset
OCT angiography	Plexus segmentation, scan size, projection artifact handling, flow metrics	OCTA dataset, optical coherence tomography angiography, capillary plexus dataset, FAZ segmentation
Fundus fluorescein angiography	Phase timing, leakage labels, vessel or lesion annotations, video vs still frames	fluorescein angiography dataset, FFA dataset, retinal angiography images, leakage segmentation
Indocyanine green angiography	Phase timing, choroidal disease labels, still frames vs sequence	ICGA dataset, indocyanine green angiography dataset, choroidal angiography
Slit-lamp photography	Illumination type, magnification, eye region, diagnosis labels	slit lamp image dataset, anterior segment image dataset, ocular surface image dataset
Anterior segment OCT	Angle metrics, corneal scans, lens or iris visibility, acquisition protocol	AS-OCT dataset, anterior segment OCT images, angle closure OCT dataset
Specular or confocal microscopy	Cell density labels, endothelial masks, magnification, device model	specular microscopy dataset, corneal endothelial cell dataset, confocal microscopy cornea dataset
Topography and tomography	Map type, raw measurements vs rendered maps, keratoconus labels, device model	corneal topography dataset, Pentacam dataset, Scheimpflug dataset, keratoconus tomography
Ultrasound and UBM	Probe type, B-scan vs UBM, lesion labels, measurement annotations	ocular ultrasound dataset, eye B-scan ultrasound dataset, UBM dataset
Visual fields and perimetry	Program type, reliability indices, pointwise thresholds, glaucoma labels	Humphrey visual field dataset, perimetry dataset, visual field progression dataset
External eye photography and video	Capture device, pose, illumination, laterality, privacy handling	external eye image dataset, eye photo dataset, periocular image dataset, mobile eye dataset
Eye tracking and gaze sensors	Sampling rate, calibration method, screen task, event labels	eye tracking dataset, gaze dataset, fixation saccade dataset, pupillometry dataset
Clinical tabular, EHR, and multimodal cohorts	Linked imaging, visit structure, de-identification, variable dictionary	ophthalmology EHR dataset, multimodal ophthalmology dataset, retinal imaging clinical dataset

Task Search Map

Task terms are often the fastest way to find annotated datasets. Combine them with anatomical and modality terms.

Task family	Typical annotations	Search terms
Disease classification	Image-level diagnosis, referable disease labels, severity grades	classification, diagnosis, screening, grading, referable, disease detection
Lesion detection	Bounding boxes, points, pixel masks, image-level lesion presence	lesion detection, microaneurysm detection, exudate detection, hemorrhage detection
Segmentation	Pixel masks, layer boundaries, cup-disc masks, vessel maps, fluid masks	segmentation, semantic segmentation, instance segmentation, boundary annotation
Grading and staging	Ordinal disease severity, clinical scales, consensus grades	severity grading, staging, diabetic retinopathy grade, cataract grade, glaucoma grade
Measurement and biomarker extraction	Thickness, cup-to-disc ratio, FAZ area, vessel caliber, cell density	biomarker dataset, thickness measurement, cup disc ratio, FAZ area, endothelial cell density
Registration and longitudinal analysis	Paired visits, image pairs, aligned modalities, progression labels	longitudinal, progression, registration, follow-up, time series
Image quality and artifact detection	Quality scores, gradability labels, artifact categories	image quality dataset, gradability, artifact detection, ungradable fundus
Domain adaptation and generalization	Multi-center, multi-device, geography, demographic metadata	multi-center, cross-device, domain adaptation, external validation
Multimodal learning	Linked fundus, OCT, visual field, tabular, text, or genetics	multimodal ophthalmology dataset, fundus OCT dataset, imaging clinical dataset
Report generation and language	Reports, captions, clinical notes, structured findings	ophthalmology report dataset, retinal report generation, image captioning eye dataset

Common Dataset Types

Distinguish the dataset type before deciding how to register it.

Dataset type	Register as	Notes
Primary public dataset	New dataset if not already covered	Prefer the official source page, data paper, and original citation.
Challenge dataset	Dataset or benchmark, depending on source structure	Record challenge access terms, train/test visibility, and whether labels are public.
Hospital or screening cohort	Dataset if data are accessible or clearly documented	Capture institution, country, cohort criteria, and access restrictions.
Data paper archive	Dataset if files are deposited and reusable terms are available	Use repository metadata and the paper for sample counts and methods.
Annotation extension	Separate dataset only if annotations are independently published and useful	Link to the base dataset and describe what annotation layer is new.
Curated subset	Usually not a new dataset unless separately released with stable access and citation	Note selection criteria and parent dataset relationship.
Mirror or rehost	Usually not a new dataset	Prefer the original source unless the mirror has distinct terms or stable derived content.
Composite benchmark	Register carefully as composite or derived	Identify every parent dataset and avoid double-counting images.
Synthetic dataset	Dataset if source generation method, files, and license are clear	Mark synthetic status and avoid presenting it as patient-derived data.
Teaching atlas or image gallery	Usually exclude unless downloadable and licensed for dataset use	Many atlases are useful references but not reusable datasets.

Qualification Criteria

Use these checks before proposing a dataset entry.

Criterion	Accept when	Reject or defer when
Public source	There is a stable source page, repository, DOI, challenge page, or institutional page	Only a secondary blog, copied archive, or broken link is available
Access path	Files are downloadable, gated with documented terms, or manually requestable	The dataset is only mentioned in a paper with no data access path
License or terms	License, terms of use, challenge rules, or data-use agreement are visible	Terms are absent, contradictory, or copied from an unofficial mirror
Ophthalmic relevance	The dataset contains eye, ocular, visual function, or ophthalmic clinical data	Eye content is incidental or not separable from a broader non-eye dataset
Metadata sufficiency	Source supports modality, task, sample unit, rough scale, and access notes	Basic identity, modality, or source provenance cannot be verified
Citation path	There is a paper, DOI, repository citation, or source-recommended citation	No stable citation or attribution path exists
Ethical and privacy status	Source describes de-identification, consent, review, or public release basis when relevant	Identifiable patient media are exposed without clear release basis
Catalog fit	It adds a new source, modality, task, annotation layer, or access route	It is only a duplicate mirror or undocumented repost

Minimum useful notes for a candidate:

• Dataset name and known aliases
• Primary URL and backup URL if relevant
• Citation or DOI
• Modality and anatomical scope
• Task and annotation type
• Sample unit: images, eyes, patients, visits, videos, volumes, or records
• Access type: direct, platform, gated, challenge, manual, or unavailable
• License or terms text exactly as stated by the source
• Relationship to existing catalog entries

Duplicate And Composite Detection Criteria

Many ophthalmology datasets appear under multiple names, mirrors, challenge pages, papers, Kaggle uploads, Hugging Face repos, and institutional pages. Check for duplication before adding a new entry.

Signals that two candidates are probably the same dataset:

• Same acronym, expanded name, or distinctive dataset title.
• Same DOI, arXiv paper, challenge page, repository record, or citation.
• Same authors, institution, recruitment site, and collection period.
• Same sample counts, class distributions, image dimensions, or train/test split sizes.
• Same file names, folder layout, checksums, or archive names.
• Same images visible in thumbnails, papers, README examples, or challenge documentation.
• Same license text and access instructions copied across mirrors.
• A source page explicitly says it mirrors, republishes, or converts another dataset.

Signals that a candidate may be a derived or composite dataset rather than a duplicate:

• It adds new labels, masks, grades, reports, quality scores, or splits to an existing image set.
• It merges multiple named datasets into a benchmark, often with harmonized labels or resized images.
• It converts formats, crops regions, normalizes images, or packages a model training subset from one or more parent datasets.
• It includes both original files and generated or synthetic augmentations.
• It cites parent datasets but publishes a new stable archive and citation.

Suggested handling:

• Prefer the original dataset as the primary catalog entry when mirrors do not add distinct annotations, access terms, or citation value.
• Register a derived annotation set only when the added annotation layer is independently accessible, citable, and useful.
• For composites, record parent datasets in notes and avoid implying that all samples are newly collected.
• For challenge splits, distinguish training data, validation labels, hidden test sets, and leaderboard-only evaluation.
• When unsure, leave a duplicate-risk note for human review instead of adding a confident new entry.

Catalog Overlap Notes

The current catalog intentionally keeps some derivative resources when they add a useful task, annotation layer, benchmark split, or machine-readable representation. These records should not be interpreted as independent patient cohorts unless their notes say so.

High-confidence duplicate, subset, or derivative relationships:

Catalog entries	Relationship	Handling
bidr, dr_arranged, eyepacs	bidr and dr_arranged are EyePACS-based reposts/subsets.	Keep for reproducibility and access history; do not count as independent cohorts.
aod, odir2019	AOD is an augmented ODIR-5K derivative.	Keep as an augmentation/preprocessing resource.
aptos_arcade_onh_masks, aptos2019	APTOS-derived vascular arcade and optic nerve head masks.	Keep as an annotation layer.
maples_dr, messidor1/messidor2	MAPLES-DR re-annotates MESSIDOR images with biomarker masks.	Keep as an annotation layer; verify exact source release when using image-level splits.
riga, riga_plus, messidor1/messidor2	RIGA includes a MESSIDOR sub-archive; RIGA+ is a domain-adaptation derivative.	Keep but flag partial overlap.
dme_vqa, dme_vqa_logical, idrid, e_ophtha	VQA layers derived from existing fundus datasets; logical variant extends DME VQA.	Keep as language/VQA task resources, not new image cohorts.
retinal_vessel_robustness, rite, stare, chase_db1	Robustness/artery-vein resources derive from common vessel datasets; RITE is DRIVE-based.	Keep task-specific annotation/benchmark records; track parent datasets.
cataract101_extended_labels, cataract_101	Label extension for Cataract-101 videos.	Keep as a label layer.
rfmid2, rfmid	Versioned auxiliary RFMiD release.	Keep as an extension, not a duplicate.
brset_mbrset_embeddings, brset, mbrset	Embedding representation derived from BRSET/MBRSET.	Keep as a representation layer.
fairvlmed, harvard_fairvision	Text/NPZ view of the Harvard FairVision cohort.	Keep as a derived multimodal view.
stage_task1, stage_task2, stage_task3	Same challenge collection exposed as separate task records.	Keep task records; avoid adding sample counts together.

Composite or broad derived corpora that need especially careful citation:

• amdnet23: compiled from ODIR, DR_200, Fundus Dataset, RFMiD, HRF, and ARIA.
• rao_fundus: web/public data plus RFMiD and JSIEC.
• fundus_domain_generalization: benchmark assembled from existing glaucoma/fundus segmentation datasets.
• smdg: aggregate glaucoma benchmark spanning many public sources.
• hassan_composite_retina, multieye, mm_retinal_reason, eyecare_100k, x_pcr, fundus_report_dataset, intraretinal_cystoid_fluid, lmod_cataract_1k, lmod_cataract_1k_cot, and ocular_chat_vqa: retain only as task-specific derived/composite resources and verify parent-source terms.

High-risk overlap clusters for future audit:

• mured versus rfmid and odir2019: similar multi-disease fundus scope, but local metadata do not prove reuse. Compare upstream papers, filenames, and hashes before flagging.
• messidor1, messidor2, riga, riga_plus, and maples_dr: high-risk MESSIDOR overlap cluster.
• stare, chase_db1, rite, and retinal_vessel_robustness: vessel segmentation overlap cluster. Add canonical DRIVE relationship handling if DRIVE is later reintroduced.

Search Workflow Record

The July 2026 follow-up search used the catalog plus PubMed, Consensus, web search, and platform pages. The intent was not to exhaust every possible archive again, but to stress-test the guide and identify high-confidence additions missed by the current catalog.

Search channels used:

• Consensus query: ophthalmology public dataset fundus OCT slit lamp cornea dataset year:2020-2026.
• PubMed/web queries: SLID slit-lamp image dataset anatomical segmentation lesion detection GitHub, ScLNet cornea scleral lens OCT layers segmentation dataset public, retinal branching angle detection dataset, FPRM retina psychological assessment dataset, and related exact-title queries.
• Platform/web queries: ScLNet cornea scleral lens OCT dataset GitHub data download, OCT and eye fundus dataset diabetic macular edema diabetic retinopathy GitHub, RETA retinal vascular tree analysis Figshare, MSHF fundus image quality assessment Figshare, OphthalVQA Figshare, OphthalWeChat Figshare, and exact repository-title searches.
• Platform API and source-page checks: GitHub repository contents and license metadata, Zenodo record metadata, Figshare DOI landing pages, PubMed/PMC data availability statements, and DOI-backed Synapse landing pages.
• Local catalog checks: eyehub search --json, direct metadata keyword scans for slid, slit, sclnet, scleral, mured, rfmid, odir, aptos, dme_vqa, riga, messidor, stare, chase, reta, mshf, ophthal, rbad, eed, and fprm.

Candidate outcomes from this pass:

Candidate	Evidence	Catalog action
SLID slit-lamp dataset	PubMed/Consensus and GitHub availability were found, but slid is already registered.	No new entry.
ScLNet cornea with scleral-lens OCT	PubMed/Consensus describes a public dataset, but the visible GitHub repository exposed code and two example images rather than the full dataset archive during review.	Defer until a stable full-data access route is verified.
SLIT-Net microbial keratitis slit-lamp data	GitHub README points to a Duke Box dataset download.	Candidate for future verification; check license/access terms before adding.
OCT and Eye Fundus Dataset for DME and DR	GitHub repository exposes fundus and OCT image folders plus CSV labels, with 1,548 fundus images and 1,113 OCT images. No repository license was declared.	Added as oct_fundus_dme_dr_mexico with Unknown license caveat.
RBAD retinal branching-angle detection	Public GitHub repository exposes a small benchmark and non-commercial academic license terms.	Added as rbad; reuse is marked research-only/non-commercial.
EED-Astig	Zenodo record describes pediatric external-eye photographs, corneal masks, keypoints, and tabular parameters under a data-use agreement.	Added as eed_astig; access remains manual/DUA.
FPRM multimodal eye imaging and psychological assessment	Scientific Data/PubMed record and Synapse DOI document fundus photographs, multispectral/functional retinal imaging, videos, labels, and psychological assessments under a DUA.	Added as fprm_retina; access remains manual/DUA.
RETA retinal vascular-tree benchmark	Figshare DOI and data paper describe IDRiD-derived images with vessel, artery/vein, bifurcation, tree, and abnormality annotations under CC BY 4.0.	Added as reta_benchmark; notes mark it as an IDRiD-derived annotation benchmark.
MSHF fundus image-quality assessment	Figshare DOI identifies a multi-source heterogeneous fundus image-quality dataset under CC BY 4.0.	Added as mshf; notes flag multi-source composition.
OphthalVQA	Figshare DOI provides a VQA benchmark dataset under CC BY 4.0.	Added as ophthalvqa; notes recommend provenance and leakage checks before validation.
OphthalWeChat	Figshare DOI exposes WeChat article and image-link metadata, not redistributed image files, with non-commercial linked-content restrictions.	Added as ophthalwechat; modality is text/index metadata and license is non-commercial.

Search Query Patterns

Start broad, then constrain by anatomy, modality, task, and source type.

"ophthalmology dataset" +"fundus"
+"retinal image dataset" +"segmentation"
+"OCT dataset" +"macular" +"fluid"
+"glaucoma dataset" +"optic disc" +"visual field"
+"corneal dataset" +"keratoconus" +"topography"
+"slit lamp dataset" +"cataract grading"
+"OCTA dataset" +"FAZ segmentation"
+"fluorescein angiography dataset" +"retina"
+"eye tracking dataset" +"gaze estimation"
+"ophthalmology EHR dataset" +"imaging"

Repository-specific patterns:

site:physionet.org ophthalmology dataset
site:zenodo.org retina dataset
site:figshare.com ophthalmology dataset
site:mendeley.com "retinal" "dataset"
site:kaggle.com "fundus" "dataset"
site:huggingface.co/datasets ophthalmology
site:grand-challenge.org retina challenge
site:codalab.lisn.upsaclay.fr ophthalmology challenge
site:synapse.org retina dataset
site:github.com "ophthalmology dataset"

Paper-oriented patterns:

"a dataset for" ophthalmology images
"publicly available" "retinal" "dataset"
"benchmark dataset" "fundus"
"data descriptor" ophthalmology
"Scientific Data" ophthalmology dataset
"segmentation dataset" "optical coherence tomography"

Access and license patterns:

"ophthalmology dataset" "license"
"retinal dataset" "CC BY"
"fundus dataset" "data use agreement"
"OCT dataset" "challenge rules"
"eye dataset" "terms of use"

Agent Notes

When an agent explores candidates, produce evidence rather than conclusions. A useful handoff note includes:

• Source URL and date accessed.
• Exact source language for license or terms.
• Evidence for sample counts and annotation types.
• Known aliases and mirror URLs.
• Duplicate or composite concerns.
• Proposed catalog action: add, defer, exclude, merge with existing entry, or request human review.

Do not download large archives during discovery. Do not bypass forms, logins, challenge registrations, data-use agreements, or institutional access controls.

PR Handoff For New Datasets

If a candidate passes qualification and is not already represented in the catalog, prepare a focused PR rather than mixing it with unrelated cleanup. For agent-driven changes, use .claude/skills/add-dataset/SKILL.md and the agent-fillable PR template in that skill directory.

Required PR evidence:

• Primary source URL, access route, and date accessed.
• Exact license or terms text. If absent, set license="Unknown" and explain the missing source terms in notes.
• Citation, DOI, PubMed ID, or repository-recommended citation.
• Modality, anatomical scope, task, sample unit, and source-reported count.
• Duplicate/composite assessment with parent datasets named explicitly.
• Files changed, generated docs status, and smoke-test output.

Do not open an add-dataset PR when the only evidence is a paper abstract, a model-code repository without data, a mirror with unclear provenance, or a dataset whose license conflicts with the original source.