MaybeRichard/zju-eye-pretrain

ZJU Eye-Pretrain Dataset

Unified multi-source ophthalmological imaging dataset for foundation model pretraining and downstream tasks.

1.18M images spanning 57 cohorts with a strict 41-column unified manifest schema.

Composition

1,179,458 images · 57 cohorts · 8 modalities, under one strict 41-column unified manifest schema. Stored as HF parquet shards in data/{private_topcon, public_fundus, public_oct, public_new}/; loadable by batch or by single cohort (57 configs, see YAML above).

modality	images	cohorts
OCT B-scan (oct_bscan)	839,729	22
color fundus (fundus_color)	284,685	28
SLO (slo_gray)	36,027	2
ultra-widefield fundus (uwf_fundus)	13,321	4
OCT-angiography en-face (octa_enface)	1,500	1
en-face OCT (oct_enface)	1,500	1
fluorescein angiography (fluorescein_angiography)	1,376	1
slit-lamp (slit_lamp)	1,320	1

Two cohorts are multi-modal and are counted under each modality they contribute: shanghai_topcon (private; OCT + SLO + fundus) and gamma_multimodal (OCT + fundus). Per-modality cohort counts therefore sum to 60 over 57 unique cohorts.

See DATASET_OVERVIEW.md for per-cohort devices, regions, masks and demographics.

OCT B-scan — 839,729 (22 cohorts)

cohort	images		cohort	images
shanghai_topcon (private)	352,343		srinivasan_2014	3,231
octa500	119,600		aroi	3,072
kermany	109,309		amd_sd	3,049
olives	63,489		rvome_oct	3,012
gamma_multimodal	51,200		mmrdr_oct	2,936
areds2	38,382		octdl	1,877
uestc	35,280		chiu_dme_2015	610
c8	24,000		octid	572
neh_ut_2021	16,810		thoct1800	96
retouch	6,936		glaucoma	49
oimhs	3,859		sparsity_sdoct_2012	17

Color fundus — 284,685 (28 cohorts)

cohort	images		cohort	images
eyepacs_combo_dr_aug	143,668		g1020_glaucoma	1,020
mfiddr_dr	34,452		fives_vessel	800
shanghai_topcon (private)	30,714		acrima_glaucoma	705
brset_multidisease	16,266		origa_glaucoma	650
ddr_dr	12,522		idrid	597
mmrdr_cfp	11,112		papila_glaucoma	488
airogs_glaucoma	9,540		rimone_glaucoma	485
odir_multidisease	6,392		adam_amd	400
lag_glaucoma	4,854		gamma_multimodal	200
rfmid_multidisease	3,200		drishti_glaucoma	101
fgadr_dr	1,842		hrf_vessel	45
messidor2_dr	1,744		stare_vessel	40
deepdrid_dr	1,600		chasedb1_vessel	28
refuge2_disc_cup	1,200		drive_vessel	20

Ultra-widefield fundus — 13,321 (4 cohorts)

cohort	images	task
mmrdr_uwf	10,388	DR + 7 lesions
uwf_tumor	2,031	retinal tumor
uwf_disease	700	7-class (AMD / DR / PM / RD / RVO / uveitis / normal)
deepdrid_uwf	202	DR grading

SLO — 36,027 (2 cohorts)

Confocal scanning-laser ophthalmoscopy, en-face grayscale (both near-infrared).

• shanghai_topcon (private) — 35,985 — Topcon DRI OCT Triton near-IR SLO localizer.
• ravir_av — 42 — Heidelberg Spectralis IR-SLO (815 nm), with artery/vein segmentation masks.

Smaller modalities (1 cohort each)

modality	images	cohort	task
OCT-angiography en-face	1,500	octa500_octa_enface	vascular maps (3 depth projections) + FAZ/vessel masks
en-face OCT	1,500	octa500_oct_enface	OCT depth projections
fluorescein angiography	1,376	iovs_fa	leakage (DME) + late-phase masks
slit-lamp	1,320	nuclear_cataract	nuclear cataract

OCT-public quality screen: B-scans with height < 256 px were removed — the nyu_poag cohort entirely (56,576) plus 1,891 sub-256px from thoct1800/octdl.

Quick Start

from datasets import load_dataset, concatenate_datasets, Image

# === Load by batch (4 batches) ===
ds_priv = load_dataset("mayberichard/zju-eye-pretrain", "private_topcon")
ds_fun  = load_dataset("mayberichard/zju-eye-pretrain", "public_fundus")
ds_oct  = load_dataset("mayberichard/zju-eye-pretrain", "public_oct")
ds_new  = load_dataset("mayberichard/zju-eye-pretrain", "public_new")   # 32 cohorts, 8 modalities

# === Load by single cohort (57 available, see configs in YAML above) ===
ds = load_dataset("mayberichard/zju-eye-pretrain", "kermany")    # 109k
ds = load_dataset("mayberichard/zju-eye-pretrain", "octa500")    # 120k

# === IMPORTANT: cast binary columns to Image for auto-decode ===
ds = ds.cast_column("image", Image())
# For cohorts with masks (DRIVE/IDRiD/REFUGE2/AROI/OIMHS/AMD-SD/Chiu/Glaucoma/OCTA500/RETOUCH):
for col in ds["train"].features:
    if col.endswith("_mask") and str(ds["train"].features[col]) == "Value('binary')":
        ds = ds.cast_column(col, Image())

# Each row after cast:
#   image: PIL.Image                                  (auto-decoded)
#   {vessel|fov|layer|lesion|disc_cup|...}_mask: PIL.Image or None
#   image_id, study_id, patient_hash, modality, anatomy, severity, diagnosis_group, ...

# === Concat all 4 batches manually if needed ===
# Note: schemas differ across batches (mask column sets), so use only shared cols:
shared = ["image_id", "cohort", "study_id", "patient_hash", "modality",
          "anatomy", "device_vendor", "device_model", "severity",
          "diagnosis_group", "image", "bscan_index"]
all_ds = concatenate_datasets([
    ds_priv["train"].select_columns(shared).cast_column("image", Image()),
    ds_fun["train"].select_columns(shared).cast_column("image", Image()),
    ds_oct["train"].select_columns(shared).cast_column("image", Image()),
    ds_new["train"].select_columns(shared).cast_column("image", Image()),
])
# 1.18M images total, mask cols dropped (use per-batch load if you need masks)

# === Streaming for big training runs (avoids downloading all ~340 GB) ===
ds = load_dataset("mayberichard/zju-eye-pretrain", "public_oct", streaming=True)
ds = ds.cast_column("image", Image())
for row in ds["train"]:
    img = row["image"]   # PIL Image, lazy-decoded
    ...

Schema (41-column manifest, identical across all batches)

cohort, study_id, patient_hash, visit_date, eye,
device_vendor, device_model, device_serial_hash, device_software_version,
hospital_domain, ethnicity,
image_quality_score, image_quality_band,
diagnosis_group, lesion_tags, lesion_location, layer_involvement, severity,
diagnosis_source, label_confidence, schema_version,
image_id, file_path, file_format,
modality, anatomy, device_technology, scan_protocol,
scan_x_mm, bscan_index,
image_height_px, image_width_px, axial_resolution_um,
has_segmentation, n_layers_visible,
fovea_x_norm, crt_um, choroid_thickness_um,
oct_footprint_bbox_fundus, oct_footprint_bbox_slo,
is_valid

Plus per-image image bytes and per-cohort mask columns.

Captions (v2 — disease/lesion-centric)

Captions were redesigned (2026-06, replacing v1) to be disease/lesion-centric and EyeDiff-style, after early training showed the old metadata-heavy captions hurt prompt-following. Text structure (comma-separated):

{modality}, {region}, {diagnosis[+severity]}, {lesions...}

Each image has 1–3 captions at increasing specificity (level column), de-duplicated:

level	example
short	OCT B-scan, diabetic macular edema
medium	color fundus, moderate non-proliferative diabetic retinopathy
dense	OCT B-scan, macula, neovascular age-related macular degeneration, choroidal neovascularization

• Disease/severity/lesion terms are standardized through a controlled vocabulary (code/disease_dict.py).
• Acquisition metadata (device, dataset name, quality score, slice index, bbox, µm thickness, eye) is kept out of the prompt — it lives in the image manifest.
• Training tip: sample one tier per image per step, and replace the caption with "" ~10% of the time for classifier-free guidance (dropout is a loader concern; empty captions are not stored).
• Private Topcon OCT currently carries only modality, region (no disease label yet) — disease pseudo-labels are pending.

Files in captions/: captions_v2.parquet (private), public_captions_v2.parquet, oct_public_captions_v2.parquet, public_new_captions_v2.parquet, plus per-cohort *_captions_v2.parquet. Total ~2.25M caption rows (100% image coverage; short on every image, medium/dense conditional by modality). Join on image_id.

import pandas as pd
caps = pd.read_parquet("hf://datasets/MaybeRichard/zju-eye-pretrain/captions/oct_public_captions_v2.parquet")
# join on image_id with the images config

Denoising / Super-resolution test set (separate — NOT part of pretraining)

A standalone paired OCT denoising test set lives under denoising_test/, kept separate from the pretraining cohorts (its own config, split=test). It is 1,735 paired 640×640 grayscale OCT B-scans (noisy input ↔ clean target; original .tif bytes). Do not mix it into the pretraining configs.

column	type	note
id	int	pair id (1–1735)
noisy	bytes	noisy OCT B-scan (.tif)
clean	bytes	clean ground-truth (.tif)
noisy_filename / clean_filename	str	original names

from datasets import load_dataset
from datasets import Image
ds = load_dataset("MaybeRichard/zju-eye-pretrain", "denoising_test", split="test")
ds = ds.cast_column("noisy", Image()).cast_column("clean", Image())  # PIL decodes the TIFF bytes

Licensing

This dataset aggregates multiple sources with mixed licenses. See LICENSE for per-cohort license terms. Users are responsible for compliance with the original license of each cohort.

Private Shanghai Topcon data is included for research convenience. Commercial use is prohibited.

Citation

If you use this dataset, please cite the original source for each cohort used (see DATASET_OVERVIEW.md).

Versioning & Updates

This dataset supports incremental updates. New cohorts can be added without touching existing data via additional shards in data/<batch>/. Captions were upgraded to v2 (disease/lesion-centric) in 2026-06; the old v1 caption files were replaced.

Quality screening (2026-06): OCT-public images with width or height < 256 px were removed (58,467 imgs): the nyu_poag cohort entirely (56,576, all 64×128), most of thoct1800 (1,704, height 149) and 187 from octdl. A full pure-black scan found 0 (low-mean "dark" frames are normal OCT and were kept). OCT-public is now 18 cohorts / 430,238 images (was 19 / 488,705); private and fundus unchanged.