From: Potential applications of artificial intelligence in image analysis in cornea diseases: a review
Year | Authors | Imaging modality | Sample size (eyes) | Study population | Outcome measures | AI algorithms | Diagnostic performance | Validation model |
---|---|---|---|---|---|---|---|---|
Corneal transplants | ||||||||
 2023 | Hayashi et al. [104] | AS-OCT | 300 | Patients undergoing DMEK | Predict graft detachment and rebubbling | EfficientNet | AUC: 0.875, Sens: 78.9% Spec: 78.6% | Hold-out validation |
 2023 | Patefield et al. [105] | AS-OCT | 24 | Patients undergoing DMEK | Predict graft detachment | ResNet-101 | Acc: 77%, Precis: 67% Spec: 45%, Sens: 92% AUC: 0.63 | Hold-out validation |
 2022 | Bitton et al. [141] | AS-OCT | 290 | Healthy and FECD pre-DMEK eyes | Corneal edema detection | U-Net models | AUC: 0.97 | N.A |
 2022 | Mujizer et al. [142] | 91 different parameters | 3647 | Patients undergoing PLK | Predict graft detachment | Logistic regression, CTA and RF | AUC: 0.65–0.72 | Cross validation |
 2021 | Hayashi et al. [106] | AS-OCT | 46 | Patients undergoing DALK | Predicting success of big bubble formation | VGG16 CNN | AUC: 0.746 | Cross validation |
 2020 | Yousefi et al. [99] | CASIA AS-OCT | 3162 | Post-surgery of PKP, LKP, DALK, DSAEK or DMEK | Predicting the need for future keratoplasty surgery | Unsupervised machine learning | No validation of prediction | N.A |
 2020 | Hayashi et al. [143] | AS-OCT | 31 | Patients post-DMEK requiring and not requiring rebubbling | Predicting the need for rebubbling post-DMEK | Multiple CNNs | AUC: 0.964, Sens: 96.7% Spec: 91.5% | N.A |
 2020 | Heslinga et al. [103] | AS-OCT | 1280 | Patients post-DMEK | Localize and quantify graft detachment | CNN | Dice: 0.896 | Hold-out validation |
 2019 | Treder et al. [144] | AS-OCT | 1172 | Patients post-DMEK | Detect graft detachment | Classifier tree | Acc: 96%, Sens: 98% Spec: 94% | Hold-out validation |