Table 10.
AI Tools for Diagnosis of Refractive Errors
| Index | Ref | Model and Explainability | Function | No. Training Samples | Validation Data | Validation Method | Evaluation Parameters |
|---|---|---|---|---|---|---|---|
| 1 | Sogawa et al51 | DL based on VGG-16 architecture (B) | Diagnosing those with and without myopic macular lesions | 910 | 910 | Cross Validation | OCT (A: 97%, Se: 90.6, Sp: 94.2) |
| 2 | Hemelings et al52 | CNN (B) | Classify pathological myopia and perform semantic segmentation of myopia-induced lesions. | 2000 | 400 | Hold-Out Validation | PM: (A:98.6%); RD: (A: 99.0%) |
| 3 | Wan et al53 | CNN (B) | Diagnosis algorithm to determine the risk of high myopia | 758 | 100 | Hold-Out Validation | A: 98.2% |
| 4 | Wu et al54 | DL (B) | Diagnosing vision-threatening conditions in high myopia | 1483 | 370 | Hold-Out Validation | Atrophy (A: 92.4%); Tractional: A: 85.3%); Neo-vascular: A: 94.2%) |
| 5 | Tong et al55 | RF (B), SVM (B), XGBoost (B), AdaBoost (B) | Classification of myopia and identification of influencing factors | 5,067 | 2,172 | Hold-Out Validation (70:30 split) | Overall: (AUC: 0.752); Primary: (Se: 41.3, Sp: 69.7, AUC: 0.710); Senior High: Se: 64.3, Sp: 98.0, AUC: 0.722) |
| 6 | Manoharan et al56 | ML (B) | Assess an individual with myopia is ”at-risk” or ”low-risk” for myopia progression. | 149 | 149 | Cross Validation | Se: 89.0, Sp: 94.0, AUC: 0.89–0.90 |
| 7 | Huang et al57 | Time-Aware Long Short-Term Memory (T-LSTM) (B) | Predict spherical equivalent (SE) for children and adolescents based on historical vision records | 398336 | 49,042 | Hold-Out Validation | T-LSTM: (MAE: 0.103 ± 0.140 D) |
| 8 | Ye et al58 | ResNeSt101 + Focal Loss (B) | Identifies 5 myopic maculopathies | 2,342 | 450 | Hold Out validation | Se: 92.8, Sp: 94.0, AUC: 0.927–0.974 |
| 9 | He et al59 | (Model A: DNN from ground up) (B); (Model B: Transferred Learned on ImageNet based models (B) | Classification of myopic maculopathy (MM) using optical coherence tomography (OCT) images | 2380 | 680 | Hold-Out validation, External validation | DNN model-A: (A: 95%); ImageNet based Model B: (A: 96.4) |
Abbreviations: VGG-16, Visual Geometry Group; OCT, Optical Coherence Tomography; ONH, Optic Nerve Head; RD, Retinal Detachment; MAE, Mean Absolute Error.