Table 4. Performance of four DL algorithms trained by three preprocessing methods in the test set.
| Item | NPRLs | ||
|---|---|---|---|
| Sensitivity (95% CI), % | Specificity (95% CI), % | Accuracy (95% CI), % | |
| Method 1 | |||
| Inception V3 | 90.4 (85.5–95.3) | 96.0 (94.4–97.6) | 94.8 (93.2–96.4) |
| ResNet50 | 37.2 (24.8–49.6) | 97.5 (96.2–98.8) | 84.9 (82.1–87.7) |
| InceptionResNetV2 | 98.1 (95.9–100) | 97.0 (95.6–98.4) | 97.2 (96.0–98.4) |
| VGG16 | 95.5 (92.2–98.8) | 94.8 (93.0–96.6) | 94.9 (93.3–96.5) |
| Method 2 | |||
| Inception V3 | 98.7 (96.9–100) | 98.7 (97.8–99.6) | 98.7 (97.9–99.5) |
| ResNet50 | 96.8 (94.0–99.6) | 97.6 (96.4–98.8) | 97.5 (96.4–98.6) |
| InceptionResNetV2 | 98.7 (96.9–100) | 99.2 (98.5–99.9) | 99.1 (98.4–99.8) |
| VGG16 | 97.4 (94.9–99.9) | 98.8 (97.9–99.7) | 98.5 (97.6–99.4) |
| Method 3 | |||
| Inception V3 | 84.0 (77.7–90.3) | 98.5 (97.5–99.5) | 95.5 (94.0–97.0) |
| ResNet50 | 71.8 (63.5–80.1) | 98.1 (97.0–99.2) | 92.7 (90.8–94.6) |
| InceptionResNetV2 | 93.6 (89.6–97.6) | 98.8 (97.9–99.7) | 97.7 (96.6–98.8) |
| VGG16 | 92.9 (88.7–97.1) | 97.3 (96.0–98.6) | 96.4 (95.0–97.8) |
Method 1, training based on original images; Method 2, training based on augmented original images; Method 3, training based on augmented histogram-equalized images. DL, deep learning system; NPRLs, notable peripheral retinal lesions, defined as the presence of lattice degeneration and/or retinal breaks; CI, confidence interval.