DeepDRiD - Data

Data

In our database, DeepDR (Deep Diabetic Retinopathy), the fundus images composed of two parts. One part is from the participants in the Shanghai Diabetic Complication Screening Project (SDCSP), Nicheng Diabetes Screening Project (NDSP), and Nationwide Screening for Complications of Diabetes (NSCD) between 2014 and 2017 for regular fundus images. The other part was captured by a retinal specialist at an Eye Clinic located in Department of Ophthalmology, Shanghai Jiao Tong University affiliated Sixth People’s Hospital, China between 2014 to 2017 for regular fundus images and between Jan 2019 to present for ultra-widefield retinal images. From the thousands of examinations available, we have extracted 2000 regular fundus images from 500 patients and 256 ultra-widefield images from another 128 patients to form our dataset.

In particular, each of the 500 patients provides four regular fundus images from two eyes. Each eye offers two images: one with the optic disc as the center and another one with the fovea as the center. For the ultra-widefield images, each of the 128 patients provides two images from two eyes. In short, we provide 628 patients data including 2256 images in total.

Experts verified that all images are of adequate quality, clinically relevant, that no image is duplicated and that a reasonable mixture of disease stratification representative of DR is present.

Fundus Camera Specifications

[1] Regular retinal fundus images were acquired using a digital fundus camera made by TOPCON, and all are centered either on the macula or optic disc. The images have a resolution of 1956×1934 pixels and are stored in jpg file format. The size of each image is about 778 KB.

[2] Ultra-widefield retinal images were acquired using Optomap P200Tx (Optos, Dunfermline, UK), and all are centered on the optic disc. The images have a resolution of 3900×3072 pixels and are stored in jpg file format. The size of each image is about 2 MB.