Table 7. Performance ratios of all the CNNs evaluated. The sensitivity and specificity were calculated using the Youden index previously described. The networks were evaluated with the total test set (370 normal and 124 glaucoma). The best option for each metric is highlighted in bold.
| Architecture | AUC | Sensitivity (%) | Specificity (%) | B-Accuracy |
|---|---|---|---|---|
| VGG19 TL | 0.9420 | 87.01 | 89.01 | 88.05 |
| VGG19 | 0.8971 | 82.26 | 86.81 | 84.53 |
| RESNET TL | 0.9300 | 91.94 | 80.00 | 85.97 |
| RESNET | 0.9193 | 83.87 | 86.15 | 85.01 |
| GOOGLENET TL | 0.8994 | 83.96 | 81.76 | 82.41 |
| GOOGLENET | 0.9269 | 84.68 | 87.25 | 85.97 |
| DENET DISC TL [42] | 0.9142 | 78.22 | 90.99 | 84.61 |
| DENET DISC [42] | 0.8371 | 79.03 | 75.16 | 77.09 |
| STANDARD CNN | 0.8969 | 79.03 | 87.03 | 83.03 |