Kaphathy/OphGen-Private

OphGen — curated institutional ophthalmic datasets

Curated copies of 11 institutional ophthalmic imaging datasets, prepared for the OphGen ophthalmic generative-model project: images as stored in the OphGen image store (de-identified, de-duplicated, native resolution), the train / validation / test split used in the project, six captions per image (short / medium / long × English / Chinese) and all original task annotations.

Private hospital data: access is restricted to the OphGen project; do not redistribute or re-upload. Images were de-identified (burned-in personal text removed); source_paths are paths in the hospital archives.

Loading

from datasets import load_dataset
ds = load_dataset("Kaphathy/OphGen-Private", "tongren16G95DFundus")            # one dataset, its own splits and labels
side = load_dataset("Kaphathy/OphGen-Private", "tongren16G95DFundus-sidecar")  # its other annotation files (path, size, bytes), if any
pre = load_dataset("Kaphathy/OphGen-Private", "all")               # every dataset (default config), e.g. for pretraining

In all, an image released by several datasets appears once per dataset (same image_id, same split); deduplicate on image_id if needed.

Datasets

Config	Dataset	Modalities	Train	Validation	Test	Tasks
BUltrasound	BUltrasound	Conventional ocular B-scan ultrasonography	101,473	5,689	5,574	0
chengdu_ffa	chengdu_ffa	FFA	3,519	473	934	1
eyFRDCls	eyFRDCls	CFP	418	53	139	1
FQ_Datasets	FQ_Datasets	CFP	2,681	402	802	1
handan_ffa	handan_ffa	FFA, CFP	26,560	1,478	1,471	0
handan_ffa2	handan_ffa2	FFA	17,465	980	991	0
tongren16G95DFundus	tongren16G95DFundus	CFP	172,345	24,419	49,171	1
tongrenHandan	tongrenHandan	CFP	7,012	987	1,974	1
trhd3_FRD_Cls66_Train	trhd3_FRD_Cls66_Train	CFP	7,780	1,106	2,176	1
yinhai_ffa	yinhai_ffa	FFA	812	34	52	0
YinhaiUBM	YinhaiUBM	UBM	75,720	4,012	3,857	0
Total			415,785	39,633	67,141

Split rules: official = official split; official_val_from_train = 10% of the official training set (by patient) used as validation; labeled_7_1_2 = patient-level 7:1:2 split with a fixed seed (no official split); pretrain_90_5_5 = 90:5:5 split for images without task labels. An image with different splits across tasks takes the strictest (test > validation > train).

Columns (dataset configs and all)

Column	Description
image_id	SHA-256 of the original file (stable identifier)
image	image bytes in the stored format; native resolution (no resizing)
split, split_rule	partition used in the project and how it was assigned
modality, modality_name	one of 13 standard ophthalmic modalities
width, height, format, field_of_view, device, eye, sex, age	metadata (empty when not documented)
representation, octa_layer, scan_area_mm	OCTA-specific fields
short_en … long_zh	captions built only from recorded facts; images in official test sets carry no task labels in captions (benchmark_holdout)
source_paths	original path(s) of the image in the hospital archive
labels	task labels of this dataset: task, name (label name, value encoding included), value, num, official_split, annotation / label_file (path of the annotation, embedded in masks or stored in the -sidecar config)
masks	raster annotation masks for this image (name matches labels.name, path matches labels.annotation)

<dataset>-sidecar configs: path, size, bytes of every other annotation file of the release (tables, JSON, NPZ, MAT, XML …); write bytes to path to restore the original layout.

Processing

Exact and pixel-identical duplicates were merged (labels kept); images with burned-in personal text or non-removable overlays, undecodable or blank files were removed; images were not resized. Where the release did not document a mask value mapping, the mapping was inferred and is stated in the label name. Full methods: OphGen supplementary methods.