Table 2.
Classification performance metrics recorded by the proposed system and original pretrained CNNs; entries with bold font are for the best performance classifier; underlined entries are for the pretrained CNN achieving the highest performance only over the other pretrained CNNs.
| Model | Size of RF1 | Number of Features in RF1 | Number of Features in RF2 | Specificity % | Sensitivity % | Accuracy % |
|---|---|---|---|---|---|---|
| Proposed DFTSA-Net DME diagnosis system | ||||||
| M1 | ¾ GFS | 4590 | 2295 | 95.5 | 97.5 | 96.8 |
| M2 | ½ GFS | 3105 | 1553 | 95.4 | 96.8 | 96.2 |
| M3 | ¼ GFS | 1553 | 777 | 95.0 | 95.5 | 95.3 |
| Pretrained CNNs | ||||||
| - | GoogLeNet | - | - | 85.8 | 95.5 | 90.3 |
| - | SqueezeNet | - | - | 73.1 | 97.1 | 85.1 |
| - | Inception-v3 | - | - | 91.1 | 97.1 | 94.0 |
| - | ResNet-50 | - | - | 62.7 | 92.5 | 77.6 |