Skip to main content
. 2022 May 16;9:839088. doi: 10.3389/fmed.2022.839088

TABLE 5.

Several state-of-the-art deep learning models for DR classification.

DL Systems Algorithm Training data set Test data set AUC Sensitivity Specificity Aim of detection
Abràmoff et al. (10) AlexNet/VGGNet Messidor-2 10 primary care practice sites from the United States NA 87 91 mtmDR
Gulshan et al. (17) Inception-V3 EyePACS, Messidor-2 Messidor-2 0.99 87 99 referable DR, operating cut point with high specificity
96 94 referable DR, operating cut point with high sensitivity
EyePACS-1 0.99 90 98 referable DR, operating cut point with high specificity
98 93 referable DR, operating cut point with high sensitivity
Zhang et al. (27) Inception-V3 SAMS, SPPH SAMS, SPPH 0.98 98 98 referable DR
Gulshan et al. (20) Inception-V4 EyePACS, Messidor-2 Aravind 0.96 90 92 referable DR
Sankara 0.98 92 95 referable DR
Bellemo et al. (28) VGGNet/ResNet SiDRP 2010-2013 Zambia mobile screening 0.97 92 89 referable DR
0.98 92 95 STDR
Sayres et al. (29) Inception-V4 EyePACS, 3 eye hospitals of India EyePACS2 0.88 92 95 referable DR

DL = deep learning, mtmDR = more than mild DR (ETDRS level 35 or higher and/or DNE), NA = not available, SAMS = The Sichuan Academy of Medical Sciences, SiDRP = Singapore Integrated Diabetic Retinopathy Screening program, SPPH = Sichuan Provincial Peoples Hospital, STDR = severe nonproliferative DR or worse.