A curated benchmark dataset of ultra-widefield fundus photographs paired with ground-truth clinical reports, assembled from two public ophthalmology datasets.

Each sample consists of a fundus image filename, a structured keyword label, and a free-text clinical report written by domain experts. The dataset is intended for evaluating medical vision-language models on fundus report generation tasks.

Split	Count
DeepDRiD	203
OUWFD	219
Total	422

Column Descriptions

Column	Type	Description
`filename`	string	Image filename (e.g. `DeepDRiD_4_r1.jpg`)
`relative_path`	string	Relative path to the image from the dataset root (e.g. `images/DeepDRiD/DeepDRiD_4_r1.jpg`)
`label`	string	Structured keyword-style ground-truth label written by a clinician (e.g. `Mild NPDR, dot hemorrhage, hard exudate`)
`source`	string	Origin dataset — either `DeepDRiD` or `OUWFD`
`keywords`	string	Key clinical findings extracted from the label (same content as `label`, kept for compatibility)
`report`	string	Free-text ground-truth clinical report describing the fundus image findings

Source Data

Original images are not redistributed. Download from the official sources below and follow their licenses.

DeepDRiD

Source: github.com/deepdrdoc/DeepDRiD (CC-BY-SA-4.0)
Download: git clone https://github.com/deepdrdoc/DeepDRiD.git
Cite: Liu, R. et al. DeepDRiD: Diabetic Retinopathy—Grading and Image Quality Estimation Challenge. Patterns 3, 100512 (2022).

OUWFD

Source: Figshare (~8.4 GB)
Download: Open the Figshare page and click Download.
Cite: He, S. et al. Open ultrawidefield fundus image dataset with disease diagnosis and clinical image quality assessment. Sci Data 11, 1251 (2024).

Downloads last month: 49

Total file size:

51.7 kB