TABLE 12.
Comparison with the existing deep learning-based state-of-the-art methods of glaucoma classification.
| Author | Year | Method and classifier | Database | No. of classes | Accuracy (%) |
|---|---|---|---|---|---|
| Akbar et al. (2022) | 2022 | Transfer learning and fusion | HRF, RIM-1, and ACRIMA | Two-class | 99.7, 89.3, and 99 |
| Kumar and Gupta. (2022) | 2022 | Transfer learning models | Private data | Two-class | 98.9 |
| Rehman et al. (2021) | 2021 | Pre-trained deep CNN | ACRIMA, ORIGA, RIM-ONE, | Two-class | 99.5 |
| architectures | AFIO, and HM | ||||
| Ajitha et al. (2021) | 2021 | Customized CNN | HRF, ORIGA, and Drishti-GS1 | Two-class | 93.86 |
| Olivas et al. (2021) | 2021 | CNNs (MobileNet and | Collected from ZeissOCT machine | Two-class | 90 |
| Inception V3) | at the Instituto de la Vision | ||||
| Li et al. (2019) | 2020 | Attention-based CNN | CGSA, Beijing Tongren and used | Two-class | 96.2 for LAG, |
| LAG and RIM-ONE for validation | 85.2 for RIM-ONE | ||||
| G´omez-Valverde et al. (2019) | 2019 | CNNs | ESPERANZA, Drishti and | Two-class | 88.05 |
| RIM-ONE | |||||
| Diaz-Pinto et al. (2019b) | 2019 | Pre-trained CNNs | ACRIMA, HRF, Drishti, RIM-ONE, | Two-class | 90.29 |
| and sjchoi86-HRF | |||||
| Proposed method | — | Pre-trained CNNs and classifier fusion | HVD | Three-class | 85.43 |
| Drishti | Three-class, | 90.55 | |||
| HVD + Drishti | Three-class | 85.18 | |||
| ACRIMA | Two-class | 99.57 | |||
| RIMONE | Two-class | 94.95 |
Bold values are showing the model results after applying classifier fusion operation.