Table 8. Performance ratios of all the CNNs evaluated considering a subset of the test set including the images from ESPERANZA and DRIHSTI-GS data sets set (343 normal and 45 glaucoma). The sensitivity and specificity were calculated using the Youden index previously described. The best option in each metric is highlighted in bold.
| Architecture | AUC | Sensitivity (%) | Specificity (%) | B-Accuracy |
|---|---|---|---|---|
| VGG19 TL | 0.9270 | 93.33 | 81.63 | 87.48 |
| VGG19 | 0.8491 | 71.11 | 89.21 | 80.16 |
| RESNET TL | 0.8957 | 84.44 | 85.71 | 85.08 |
| RESNET | 0.8778 | 71.11 | 93.00 | 82.06 |
| GOOGLENET TL | 0.8821 | 86.67 | 76.68 | 81.67 |
| GOOGLENET | 0.9276 | 80.00 | 90.96 | 85.48 |
| DENET DISC TL [42] | 0.9272 | 82.22 | 91.55 | 86.88 |
| DENET DISC [42] | 0.8261 | 77.78 | 75.51 | 76.64 |
| DENET [42] | 0.7507 | 70.45 | 70.26 | 70.35 |
| STANDARD CNN | 0.8312 | 75.76 | 78.72 | 77.14 |