Table 1.
A summary table of artificial intelligence (AI) application in the diagnosis of keratoconus and dry eye diseases, in reverse chronological order
| Year | Authors | Imaging modality | Sample size (eyes) | Study population | Outcome measures | AI algorithms | Diagnostic performance | Validation method |
|---|---|---|---|---|---|---|---|---|
| Keratoconus | ||||||||
| 2023 | Lu et al. [15] | Pentacam, SD-OCT, APT | 599 | Healthy, FF, early, advanced KC eyes | KC detection | RF/CNN | AUC: 0.801–0.902 | Hold-out validation |
| 2023 | Kundu et al. [111] | AS-OCT | 1125 | Healthy, VAE and KC eyes | KC detection | RF |
AUC: 0.994–0.976, Acc: 95.5%–95.6% Sens: 71.5%–98.5%, Precis: 91.2%–92.7% |
Hold-out validation |
| 2022 | Cohen et al. [112] | Galilei | 8526 | Healthy, suspect and KC eyes | KC detection | RF |
AUC: 0.964–0.969, Acc: 90.2%–91.5% Sens: 94.2%–94.7%, Spec: 89.6%–89.8% |
Hold-out validation |
| 2022 | Almeida Jr et al. [113] | Pentacam | 2893 | Healthy, VAE and KC eyes | KC detection | BESTi MLRA |
AUC: 0.91, Sens: 86.02% Spec: 83.97% |
Hold-out validation |
| 2022 | Reddy et al. [114] | Oculyzer | 1331 | Healthy and KC eyes | Prediction of latent progression of KC | CNN | 11.1 months earlier progression than KP (P < 0.001) | Hold-out validation |
| 2022 | Gao et al. [115] | Pentacam | 208 | Healthy, subclinical and KC eyes | Subclinical and KC detection | KeratoScreen ANN |
Sens: 93.9%–97.6% Precis: 95.1%–96.1% |
Hold-out validation |
| 2022 | Xu et al. [116] | Pentacam | 1108 | Healthy, VAE and KC eyes | Detection of healthy eye in VAE | KerNet CNN |
Acc: 94.67% AUC: 0.985 |
Hold-out validation |
| 2022 | Gairola et al. [117] | SmartKC | 57 | Healthy and KC eyes | KC detection | CNN |
Sens: 91.3% Spec: 94.2% |
Hold-out validation |
| 2022 | Lu et al. [65] | SD-OCT, APT | 622 | Healthy, FF, early, advanced KC eyes | KC detection | RF/CNN |
AUC: 0.99, Sens: 75% Spec: 94.74% |
Hold-out validation |
| 2022 | Subramaniam et al. [118] | Pentacam | 900 | Healthy, subclinical and KC eyes | KC detection and grading | PSO, GoogLeNet CNN |
Acc: 95.9%, Spec: 97.0% Sens: 94.1% |
Hold-out validation |
| 2022 | Mohammadpour et al. [12] |
Pentacam, Sirius, OPD-Scan III Corneal Navigator |
200 | Healthy, subclinical and KC eyes | KC detection | RF |
Subclinical KC – Acc: 88.7%, Sens: 84.6%, Spec: 90.0% KC – Acc: 91.2%, Sens: 80.0%, Spec: 96.6% (Based on Sirius Phoenix) |
N.A |
| Dry eye diseases | ||||||||
| 2023 | Shimizu et al. [54] | ASV | 158 | Healthy and DED eyes | DED grading based on TBUT | ImageNet-22 k CNN |
Acc: 78.9%, AUC: 0.877 Sens: 77.8%, Spec: 85.7% |
Hold-out validation |
| 2023 | Abdelmotaal et al. [52] | ASV | 244 | Healthy and DED eyes | DED detection | CNN | AUC: 0.98 | Hold-out validation |
| 2022 | Fineide et al. [51] | ASV | 431 | Patients with DED | DED grading based on TBUT | RF |
Sens: 99.8%, Precis: 99.8% Acc: 99.8% |
Cross validation |
| 2022 | Edorh et al. [119] | AS-OCT | 118 | Healthy and DED eyes | Epithelial changes as a marker of DED | RF |
Sens: 86.4% Spec: 91.7% |
N.A |
| 2021 | Chase et al. [44] | AS-OCT | 151 | Healthy and DED eyes | DED detection | VGG19 CNN |
Acc: 84.62%, Sens: 86.36% Spec: 82.35% |
Hold-out validation |
| 2021 | Elsawy et al. [120] | AS-OCT | 879 | Healthy and various anterior segment eye diseases | DED detection | VGG19 CNN | AUC: 0.90–0.99 | Hold-out validation |
| 2020 | Maruoka et al. [62] | HRT-3 confocal microscopy | 221 | Healthy and obstructive MGD eyes | Obstructive MGD detection | Multiple CNNs |
AUC: 0.96, Sens: 94.2% Spec: 82.1% |
Hold-out validation |
| 2020 | da Cruz et al. [56] | Doane interferometer | 106 | VOPTICAL_GCU database of tear film images | Classification of tear film lipid layer | SVM, RF, NBC, MLP, RBFNetwork, random tree |
Acc: 97.54%, AUC: 0.99 κ: 0.96 |
Cross validation |
| 2020 | Stegmann et al. [121] | AS-OCT | 6658 | Healthy eye images | Tear meniscus segmentation | TBSA CNN |
Sens: 96.4%, Spec: 99.9% Jaccard index: 93.2% |
Cross validation |
| 2019 | Wang et al. [122] | Keratograph 5 M | 209 | Healthy and DED eyes | Segmentation of meibomian gland, grading of meiboscore | CNN |
Acc: 95.4%–97.6% IoU: 66.7%–95.5% |
Hold-out validation |
| 2018 | Arita et al. [55] | DR-1α tear interferometer | 100 | Healthy and DED eyes | DED detection and grading | Interferometric movies |
κ: 0.76 Acc: 76.2%–95.4% |
N.A |
Acc = accuracy; ANN = artificial neural network; APT = air puff tonometry; AS-OCT = anterior-segment optical coherence tomography; ASV = anterior segment videography; AUC = area under curve; CNN = convolutional neural network; DED = dry eye disease; FF = forme fruste keratoconus; IoU = intersection over union; IVCM = in vivo confocal microscopy; κ = kappa index; KC = keratoconus; KP = keratometric progression; MGD = meibomian gland disease; MLP = multilayer perceptron; N.A. = not available; NBC = naïve Bayes classifier; Precis = precision; RF = random forest; SD-OCT = spectral-domain optical coherence tomography; Sens = sensitivity; Spec = specificity; SVM = support vector machines; TBSA = threshold based algorithm; TBUT = tear breakup time; VAE = very asymmetric eyes (fellow to KC eyes)
Jaacard index is a statistical analysis of how similar two sample sets are