Skip to main content

Table 1 Summary of application of AI in the screening or diagnosis of cataract

From: Application of artificial intelligence in cataract management: current and future directions

Year

Authors

Imaging

Sample size

AI algorithms

AUC

Accuracy (%)

Sensitivity (%)

Specificity (%)

Adult cataract diagnosis (screening and grading)

2020

Li et al. [10]

SLP

1772

ResNet-CNN

–

D = 98.4–99.8

D = 99.4

D = 99.1

2019

Wu et al. [11]

SLP

37,638

ResNet

D = 0.90–1.00

G = 0.86–0.97

D = 84.2–99.5

G = 73.2–94.9

D 60.1–99.5

G = 63.2–92.1

D = 76.4–99.6

G = 63.2–92.1

2019

Xu et al. [12]

FP

8030

AlexNet + VisualDN

–

D + G = 86.2

D + G = 79.8–95.0

D + G = 83.3–88.4

2019

Zhang et al. [75]

FP

1352

SVM + FCNN

–

G = 93.0

D = 99.4

G = 82.4–96.4

–

2017

Xiong et al. [76]

FP

1355

BPNN + MCDA

–

D = 92.8

G = 81.1–83.8

D = 93.1

D = 92.1

2016

Yang et al. [77]

FP

1239

Ensemble learning (SVM + BPNN)

–

D = 92.0–93.2

G = 83.9–84.5

D = 91.4–94.2

G = 62.5–79.5

D = 91.5–92.5

G = 87.9–98.9

2015

Guo et al. [78]

FP

445

MCDA

–

D = 90.9

G = 77.1

–

–

2015

Gao et al. [9]

SLP

5378

CRNN

–

G = 70.7

–

–

2013

Xu et al. [79]

SLP

5378

SVR

–

G = 69.0 (with up to 98.9 for within 1-step error)

–

–

2012

Gao et al. [80]

SLP

434

–

–

D = 62.0

–

–

2011

Cheung et al. [81]

SLP

5547

SVM

D = 0.88–0.90

–

D = 79.7–83.7

D = 79.5–81.9

2010

Acharya et al. [82]

SLP

140

BPNN

–

D = 93.3

D = 98.0

D = 100.0

Intraocular lens power calculation methods and biometry

2021

Ladas et al. [38]

Data

1391

SVR, XGB, ANN

–

PE within 0.5 D = 80.0 (SRK + SVR)

81.0 (SRK + XGB)

67.0 (SRK + ANN)

82.0 (Holladay I + SVR)

82.0 (Holladay I + XGB)

80.0 (Holladay I + ANN)

82.0 (LSF + SVR)

81.0 (LSF + XGB)

81.0 (LSF + ANN)

MAE = 0.325 (SRK + SVR)

0.314 (SRK + XGB)

0.439 (SRK + ANN)

0.307 (Holladay I + SVR)

0.309 (Holladay I + XGB)

0.326 (Holladay I + ANN)

0.311 (LSF + SVR)

0.310 (LSF + XGB)

0.319 (LSF + ANN)

–

–

2021

Debellemanière et al. (PEARL-DGS) [39]

Data

6120

SVR, GBT, RM

–

PE within 0.50 D = 87.4

MAE = 0.443 (short); 0.240 (long)

–

–

2020

Carmona et al. (Karmona) [23]

Data

260

SVM-RBF

MARS-SOP

–

PE within 0.50 D = 90.4

MAE = 0.240

–

–

2019

Connell et al. (Kane) [83]

Data

846

RM

–

PE within 0.50 D = 77.9

MAE = 0.441 (short); 0.322 (medium); 0.326 (long)

–

–

2019

Wan et al. (Hill-RBF 2.0) [19]

Data

127

RM

–

PE within 0.50 D = 86.6

–

–

2019

Sramka et al. [17]

Data

2194

SVM-RM and

MLNN-EM

–

PE within 0.50 D = 82.3–82.7

–

–

2016

Koprowski et al. [48]

Data

173

CNN

–

ECPP 0.16 ± 0.14 Dp

–

–

Intraoperative tools

2020

Lanza et al. [28]

Surgery factors

1229

DA

–

68.4

–

–

2020

Lecuyer et al. [26]

Cataract surgery videos

50

CNN (VGG19, InceptionV3, ResNet50)

–

70.0–84.4

–

–

2019

Yu et al. [24]

Cataract surgery videos

100

SVM, RNN, CNN (SqueezeNet), CNN-RNN

0.71–0.77

91.5–95.9

0.5–97.4

87.7–99.9

Postoperative assessment of posterior capsule opacification

2018

Jiang et al. [29]

SLP

6090

TempSeq-Net

0.97

92.2

81.0

91.4

2012

Mohammadi et al. [30]

SLP

352

ANN

0.71

–

25

97

Pediatric cataract assessment

2020

Lin et al. [51]

SLP

1738

RF

ADA

0.86 (RF)

0.85 (ADA)

86.0 (RF)

85.0 (ADA)

80.0 (RF)

77.0 (ADA)

91.0 (RF)

90.0 (ADA)

2019

Lin et al. [52]

SLP

350

CNN

–

D = 87.4

G = 70.8–90.6

D = 89.7

G = 84.2–91.3

D = 86.4

G = 44.4–88.9

2017

Liu et al. [53]

SLP

886

CNN

D = 0.97

G = 0.96–0.99

D = 97.1

G = 89.0–92.7

D = 96.8

G = 90.8–93.9

D = 97.3

G = 82.7–91.1

2017

Long et al. [54]

SLP

1349

CNN

D = 0.92–1.00

G = 0.96–1.00

D = 92.5–98.9

G = 84.6–100

D = 98.8–100

G = 85.7–100

D = 71.4–99.0

G = 90.5–100

2020

Long et al. [55]

HR

594

RF

D = 0.94

D = 89.4–98.1

D = 84.9–98.9

D = 86.9–99.0

  1. ADA= adaptive boost modelling; AI = artificial intelligence; ANN = artificial neural network; AUC = area under the curve; BPNN = back propagation neural network; CNN = convolutional recursive neural network; CRNN = convolutional recurrent neural network; D = diagnosis; DA = discriminant analysis; DN = deconvolutional network; Dp = diopters; ECPP = error of corneal power prediction error; EM = expectation–maximization; EN = ensemble model; FCNN = fully connected neural network; FP = fundus photography; G = grading; GBT = gradient boosted trees; LSF = line spread function; MAE = mean absolute error; MARS = multivariate adaptive regression spline; MCDA = multi-class discriminant analysis; MLNN = multilayer neural network; PE = percentage of eyes; RBF = radial basis function; RF = random forrest; RM = regression model; RNN = recurrent neural network; SLP = slit-lamp photography; SOP = second order polynomials; SRK = formula created by John Retzlaff, Kraff and Sanders; SVM = support vector machine; SVR = support vector regression; XGB = extreme gradient boost