Table 4.
Summary of cataract phase classification models.
| Dataset name | Dataset size | Method | AUPRC | AUROC | PPV | Sensitivity | Specificity | Accuracy | F1 score | |
|---|---|---|---|---|---|---|---|---|---|---|
| Touma et al. (this study) | Cataract 21 & Cataract 101 | 122 | Google AutoML Video Intelligence | 0.855 | NR | 81.0% | 77.1% | 98.0% | 96.0% | 0.79 |
| Yu et al.3 | ||||||||||
| Model #3 | Own dataset | 100 | CNN input with cross-sectional image data | NR | 0.712 | 78.6% | 74.5% | 97.5% | 95.6% | NR |
| Model #4 | Own dataset | 100 | CNN-RNN input with a time series of images | NR | 0.752 | 62.0% | 59.3% | 95.6% | 92.1% | NR |
| Primus et al.19 | Cataract 21 | 21 | GoogLeNet CNN | NR | NR | 69.0% (74.0%) | 67.0% (72.0%) | NR | NR |
0.68 (0.73) |
| Zisimopoulos et al.20 | CATARACTS | 50 | ResNET-RNN | NR | NR | NR | NR | NR | 78.3% | 0.75 |
| Quellec et al.21 | Own dataset | 186 | Adaptive spatiotemporal polynomial (real-time detection) | NR | NR | NR | NR | NR | 85.3% | NR |
| Qi et al.22 | Cataract 101 | 101 | ResNET (real-time detection) | NR | NR | NR | NR | NR | 88.1% | NR |
| Charrière et al.23 | Own dataset | 30 | Bayesian network (real-time detection) | NR | 0.828 | NR | NR | NR | NR | NR |
| Lalys et al.24 | Own dataset | 20 | Dynamic time wrapping and hidden Markov models (real-time detection) | NR | NR | NR | NR | NR | 95.0% | NR |
AUPRC Area under the precision-recall curve, AUROC Area under the receiver operating characteristic, PPV Positive predictive value.