Skip to main content
. 2023 May 4;11:1173094. doi: 10.3389/fcell.2023.1173094

TABLE 1.

Summary table of deep learning studies for glaucoma diagnosis with OCT images as input.

References Year Model Dataset Aim Result
Ran et al. 2019 ResNet GON/No GON detect GON Primary validation: AUROC: 0·969, sensitivity: 89%, specificity: 96%, accuracy: 91%
— — Training, testing, and primary validation dataset 2926/1951 — External validation: AUROC: 0·893–0·897, sensitivities: 78%–90%, specificities: 79%–86%, accuracies: 80%–86%
— — External validation dataset 1434/610 — —
Noury et al. 2022 DiagFind Glaucoma/Non-glaucoma manifest glaucoma AUC: perimetric glaucoma
Stanford: 0.91
Hong Kong: 0.80
India: 0.94
Nepal: 0.87
— — Training 1022/542 — —
— — Validation 142/61 — —
— — Test 453/241 — —
— — External validation dataset 1642/1035 — —
Ran et al. 2022 ResNet yes GON and yes MF/no GON and yes MF/yes GON and no MF/no GON and no MF GON AUROC
MF GON: Internal validation 0.949
External testing dataset 0.890–0.950
— — Training 1679/890/629/721 — MF: 0.855–0.896
— — Tuning 195/163/32/70 — —
— — Internal validation 205/114/36/99 — —
— — External testing dataset 1347/515/677/777 — —
Asaoka et al. 2019 Deep learning Glaucoma/Non-glaucoma early glaucoma AUC
Pretraining: 93.7%
Without pretraining: 76.6%–78.8%
Pretraining 1371/193 — —
Training 94/84 — —
Test 114/82 — —
Medeiros et al. 2021 ResNet50 86 123 progressive glaucomatous changes over time AUC: 0.86
Medeiros et al.
2019 ResNet34 Normal/Suspect/Glaucoma quantify glaucomatous structural damage MAE: 7.39 μm
AUC: predictions: 0.944
actual measurements: 0.940
— — Training 3982/13 410/9136 — —
— — Test 877/3345/2070 — —
Kamalipour et al. 2023 CNNA Normal/Suspect/Glaucoma estimate central 10° visual field MAE
CNNA: 4.04 dB
— CNNT Training and Validation 174/367/623 — —
— LR Test 20/71/110 — —
Christopher et al. 2021 ResNet50 10-2 Visual Field/24-2 Visual Field estimating visual function 10-2
R2 MD:0.82
PSD: 0.69
MAE MD: 1.9 dB
— — Training 2131/277 — 24-2
R2 MD:0.79
PSD: 0.68
MAE MD: 2.1 dB
— — Test 2674/325 — —
Lee et al. 2020a HDLM Normal/Suspect/Glaucoma predicts macular ganglion cell-inner plexiform layer thickness MAE: 4.76 μm
— — 292/109/388 — —
Hao et al. 2022 ResNet + LSTM Glaucoma/Non-glaucoma angle-closure screening AUC
Casia dataset: Images 0.766; Original videos 0.820; Aligned videos 0.905.
— — 159/210 — Zeiss dataset: Images 0.767; Original videos 0.837; Aligned videos 0.919
Xu et al. 2019 ResNet18 Open angle/Closed angle detect gonioscopic angle closure and primary angle closure disease AUC: gonioscopic angle: 0.928
disease: 0.952
— — Cross-validation 1632/1764 — —
— — Test 311/329 — —
Li et al. 2022b ResNet34 Task I/Task II Task I (1) narrow iridocorneal angles Task I
AUC: 0.943, sensitivity: 0.867, and specificity: 0.878
— — Training 4515/378 Task II (2) peripheral anterior synechiae Task II
AUC: 0.902, sensitivity: 0.900, and specificity: 0.890
— — Internal validation 1101/376 — —
— — External testing 2222/102 — —
Randhawa et al. 2021 ResNet18 Open angle/Closed angle detect gonioscopic angle closure AUC: 0.894–0.922
— — CHES train 1764/1632 — —
— — CHES test 329/311 — —
— — Singapore 570/9595 — —
— — USC 66/234 — —
Shon et al. 2022a β-VAE Training 1692 extract a low-dimensional latent structure mean values of visual field index: 86.4%
mean deviation: −5.33 dB
— — Validation 419 — —
Shon et al. 2022b VAE Training 1692 Analysis the latent structure Among the symmetrical latent variables, the first three and the last demonstrated easily recognized features.
— — Validation 419 — —
Muhammad et al. 2017 HDLM Glaucoma/Health or suspects Distinguish glaucoma eyes accuracy: 63.7%–93.1%
— — 57 eye/45 eye — —
Butola et al. 2020 LightOCT Choroidal neovascularization/Diabetic macular edema/Drusen/Normal Distinguish glaucoma eyes accuracy: 96%
— — Training 27 206/11 349/8617/51 140 — —
— — Test 250/250/250/250 — —
Yang et al. 2021 InceptionResNetV2 Open angle/Closed angle detect the static gonioscopic angle closure and peripheral anterior synechia static gonioscopic angle closure
AUC: 0.963 sensitivity: 0.929
specificity: 0.877
— — Training 3 4705/1 5945 — appositional from synechial angle closure
AUC: 0.873
Sensitivity: 0.846
Specificity 0.764
— — Validation 8037/3254 — —
— — Test 7860/3024 — —
Soltanian-Zadeh et al. 2021 WeakGCSeg Training samples/Testing samples Cell-level quantitative features of retinal ganglion cells WeakGCSeg is on par with or superior to human experts and is superior to other state-of-the-art networks.
— — Subject 1 (IU/IU) Healthy: 7:14/1:2 — —
— — Subject 2 (IU/IU) Healthy: 7:14/1:1 — —
— — Subject 3 (FDA/FDA) Healthy: 3:4-5/1:1-2 — —
Glaucoma: 4:8/1:2
— — Subject 4 (IU/FDA Healthy: 8:16/4:6 — —
FDA/IU Healthy: 4:6/8:16
IU + FDA/IU + FDA) Healthy: 9:16–17/9:16–17

● ResNet residual network, GON, glaucomatous optic neuropathy; AUROC, area under the receiver operating characteristic; AUC, area under curve; MF, myopic features; MAE mean absolute error; CNN, convolutional neural network; LR, ordinary least squares linear regression models; MD, mean deviation; PSD, pattern standard deviation; HDLM, hybrid deep learning method; LSTM, long short-term memory; CHES, the Chinese American Eye Study; USC, the University of Southern California; VAE variational auto-encoder; IU, the Indiana University; FDA, the U.S., food and drug administration.