Glaucoma is a multifactorial neurodegenerative illness requiring early diagnosis and strict monitoring of the disease progression. Artificial intelligence algorithms can extract various optic disc features and automatically detect glaucoma from fundus photographs.
Author | Year | N. of Images | Structure | SEN | SPEC | ACC | AUC | |||||
---|---|---|---|---|---|---|---|---|---|---|---|---|
Kolar et al. [16][12] | 2008 | 30 | FD | |||||||||
OCT Fundus | Thompson et al. | 93.80% | ||||||||||
[ | 47][43] | 2019 | 1. Global BMO-MRW prediction | ResNet34 | Nayak et al. [17][13] | 2009 | 61 | Morphological | 100% | 80% | 90% | |
Bock et al. [18][14] | 2010 | 575 | Glaucoma Risk Index | 73% | 85% | 80% | ||||||
Acharya et al. [19][15] | 2011 | 60 | SVM | 91% | ||||||||
Dua et al. [20][16] | 2012 | 60 | DWT | |||||||||
0.914 | ||||||||||||
Author | Year | Outcomes Mesures | Architecture | SEN | SPEC | ACC | AUC | ||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Asaoka et al. [69][65] | 0.945 | ||||||||||||||
2016 | Pre-perimetric VFs vs. VFs in healthy eyes | FNN | 0.926 | 2. Yes glaucoma vs. No glaucoma | |||||||||||
Kucur et al. [75] | |||||||||||||||
[71] | 2018 | Early glaucomatous VF loss vs. no glaucoma | CNN with Voronoi representation | Medeiros et al. [53][49] | 2019 | 1. RNFL thickness prediction | ResNet34 | 80% | 83.7% | ||||||
Li et al. [12][ | 0.944 | ||||||||||||||
8] | 2018 | Glaucomatous VF loss vs. no glaucoma | VGG I 5 | 93% | 83% | 88% | 0.966 | 2. Glaucoma vs. Suspect/healthy | |||||||
Li et al. [76][72] | 2018 | Glaucoma vs. Healthy | VGG | 93% | 3% | 0.966 | 93.3% | ||||||||
Jammal et al. [52][48] | 2020 | RNFL prediction | ResNet34 | 0.801 | Mookiah et al. [21][17] | 2012 | 60 | DWT, HOS | 86.7% | ||||||
Berchuck et al. Lee et al. | 93.3% | 93.3% | |||||||||||||
[62][58] | 2021 | RFNL prediction | M2M | ||||||||||||
Medeiros et al. [54][50] | 2021 | Detection of RFNL thinning from fundus photos | CNN | ||||||||||||
[77][73] | 2019 | Rates of VF progression compared to SAP MD; Prediction of future VF compared to point-wise regression predictions | Deep variational autoencoder | ||||||||||||
Wen et al. [78][74] | 2019 | HFA points and Mean Deviation | CascadeNet-5 | Noronha et al. [22][18] | 2014 | OCT 2D | Asaoka et al. [55][51] | 2019 | Early POAG vs. no POAG | ||||||
Kazemian et al. [74][70 | 272 | Higher order cumulant features | 100% | 92% | 92.6% | ||||||||||
] | 2018 | Forecasting visual field progression | Kalman Filtering Forecasting | Acharya et al. [23][19] | 2015 | 510 | Gabor transform | Novel CNN | 80% | 83.3% | |||||
Garcia et al. [73 | 0.937 | ||||||||||||||
][69] | 2019 | Forecasting visual field progression | Kalman Filtering Forecasting | 89.7% | 96.2% | 93.1% | |||||||||
Isaac et al. [24][20] | 2015 | 67 | Cropped input image after segmentation | 100% | 90% | 94.1% | |||||||||
Muhammad et al. [63][59] | 2017 | ||||||||||||||
DeRoos et al. [72] | Early glaucoma vs. health/suspected eyes | CNN + transfer learning | [ | 93.1% | 0.97 | ||||||||||
68] | 2021 | Forecasting visual field progression | Kalman Filtering Forecasting | Raja et al. [25][21] | 2015 | 158 | Hybrid PSO | 97.5% | 98.3% | 98.2% | |||||
Lee et al. [64][60] | 2020 | GON vs. No GON | CNN (NASNet) | 94.7% | 100% | 0.990 | Singh et al. [26][22] | 2016 | 63 | ||||||
Devalla et al. [65][61 | Wavelet feature extraction | ] | 100% | 90.9% | 94.7% | ||||||||||
2018 | Glaucoma vs. normal | Digital stain of RNFL | 92% | 99% | 94% | Acharya et al. [27][23] | 2017 | 702 | kNN (K = 2) Glaucoma Risk index | 96.2% | 93.7% | 95.7% | |||
Maheshwari et al. | |||||||||||||||
Wang et al. [56][52] | 2020 | Glaucoma vs. no glaucoma | CNN + transfer learning | 0.979 | [28][24] | 2017 | 488 | Variational mode decomposition | |||||||
Thompson et al. [51] | 93.6% | [47 | 95.9% | ] | 94.7% | ||||||||||
2020 | POAG vs. no glaucoma | ResNet34 | 95% | 81% | 0.96 | Raghavendra et al. [29][25] | 2017 | 1000 | RT, MCT, GIST | ||||||
97.80% | 95.8% | 97% | |||||||||||||
Pre-perimetric vs. no glaucoma | 95% | 70% | 0.92 | Ting et al. [7][3] | 2017 | 494,661 | VGGNet | 96.4% | |||||||
87.2% | 0.942 | ||||||||||||||
Glaucoma with any VF loss (perimetric) vs. no glaucoma | 95% | 80% | 0.97 | Kausu et al. [30][26] | 2018 | 86 | Wavelet feature extraction, Morphological | 98% | 97.1% | 97.7% | |||||
Mild VF loss vs. no glaucoma | 95% | 85% | 0.92 | Koh et al. [31][27] | 2018 | 2220 | Pyramid histogram of visual words and Fisher vector | 96.73% | 96.9% | 96.7% | |||||
Moderate VF loss vs. no glaucoma | 95% | 93% | 0.99 | Soltani et al. [32][28] | 2018 | 104 | Randomized Hough transform | 97.8% | 94.8% | 96.1% | |||||
Li et al. [12][8] | 2018 | 48,116 | Inception-v3 | 95.6% | 92% | 92% | 0.986 | ||||||||
Fu et al. [33][29] | 2018 | 8109 | Disc-aware ensemble network (DENet) | 85% | 84% | 84% | 0.918 | ||||||||
Raghavendra et al. [29][25] | 2018 | 1426 | Eighteen-layer CNN | 98% | 98.30% | 98% | |||||||||
Christopher et al. [34][30] | 2018 | 14,822 | VGG6, Inception-v3, ResNet50 | 84–92% | 83–93% | 0.91–0.97 | |||||||||
Chai et al. [35][31] | 2018 | 2000 | MB-NN | 92.33% | 90.9% | 91.5% | |||||||||
Severe VF loss vs. no glaucoma | 95% | 98% | 0.99 | ||||||||||||
Mariottoni et al. [66][62] | 2020 | Global RNFL thickness value | ResNet34 | ||||||||||||
OCT 3D | Ran et al. [58][54] | 2019 | Yes GON vs. No GON | CNN (NASNet) | 89% | 96% | 91% | 0.969 | |||||||
78–90% | 86% | 86% | 0.893 | ||||||||||||
Maetschke et al. [57][53] | 2019 | POAG vs. no POAG | Feature-agnostic CNN | 0.94 | |||||||||||
0.92 | Ahn et al. [36][32] | 2018 | 1542 | Inception-v3 Custom 3-layer CNN |
84.5% 87.9% |
||||||||||
Russakoff et al. [59][55] | 0.93 | 2020 | Referable glaucoma vs. non-referable glaucoma | gNet3D-CNN | 0.94 | ||||||||||
0.88 | Shibata et al. [37][33] | 2018 | 3132 | ResNet-18 | |||||||||||
AS-OCT | Fu et al. [60] | [56] | 2019 | Open angle vs. Angle closure | 0.965 | ||||||||||
VGG-16 + transfer learning | 90% | 92% | 0.96 | Mohamed et al. [38][34] | 2019 | 166 | Simple Linear Iterative Clustering (SLIC) | 97.6% | 92.3% | 98.6% | |||||
Fu et al. [67][63] | 2019 | Open angle vs. Angle closure | CNN | 0.9619 | Bajwa et al. [39][35] | 2019 | 780 | R-CNN | 71.2% | 0.874 | |||||
Xu et al. [61][57] | 2019 | 1. Open angle vs. angle closure | CNN (ResNet18) + transfer learning | 0.928 | Liu et al. [40][36] | 2019 | 241,032 | ResNet (local validation) | 96.2% | 97.7% | 0.996 | ||||
2. Yes/PACD vs. no PACD | Al-Aswad et al. [41][37] | 2019 | 110 | ResNet-50 | 83.7% | 88.2% | 0.926 | ||||||||
0.964 | |||||||||||||||
Hao et al. [68][64] | 2019 | Open angle vs. Narrowed Angle vs. Angle closure | MSRCNN | Asaoka et al. [42][38] | 2019 | 3132 | ResNet-34 | 0.965 | |||||||
ResNet-34 without augmentation | 0.905 | ||||||||||||||
VGGI I | 0.955 | ||||||||||||||
VGGI 6 | 0.964 | ||||||||||||||
Inception-v3 | 0.957 | ||||||||||||||
Kim et al. [43][39] | 2019 | 1903 | Inception-V4 | 92% | 98% | 93% | 0.99 | ||||||||
Orlando et al. [44][40] | 2019 | 1200 | Refuge Data Set | 85% | 97.6% | 0.982 | |||||||||
Phene et al. [45][41] | 2019 | 86,618 | Inception-v3 | 80% | 90.2% | 0.945 | |||||||||
Rogers et al. [46][42] | 2019 | 94 | ResNet-50 | 80.9% | 86.2% | 83.7% | 0.871 | ||||||||
Thompson et al. [47][43] | 2019 | 9282 | ResNet-34 | 0.945 | |||||||||||
Hemelings et al. [15][11] | 2020 | 8433 | ResNet-50 | 99% | 93% | 0.996 | |||||||||
Zhao et al. [48][44] | 2020 | 421 | MFPPNet | 0.90 | |||||||||||
Li et al. [49][45] | 2020 | 26,585 | ResNet101 | 96% | 93% | 94.1% | 0.992 |
Author | Year | Outcome Measures | Arch | SEN | SPEC | ACC | AUC |
---|