Table 2 The prediction results from an external validation dataset (RFMiD) to detect epiretinal membrane in fundus photographs
CNN architectures | Training set | ROC-AUC (95% CI) | Sensitivity (%, 95% CI) | Specificity (%, 95% CI) | PPV (%, 95% CI) | NPV (%, 95% CI) |
|---|---|---|---|---|---|---|
ResNet50 | Original set (no augmentation) | 0.873 (0.806–0.941) | 65.3 (44.3–82.7) | 90.4 (87.9–92.5) | 20.9 (15.5–27.6) | 98.5 (97.5–99.1) |
Original set + classic augmentationa | 0.849 (0.820–0.875) | 88.5 (69.8–97.5) | 69.5 (65.8–72.9) | 10.1 (8.6–11.9) | 99.4 (98.1–99.8) | |
Original set + DCGAN | 0.914 (0.869–0.959) | 88.4 (69.8–97.5) | 76.8 (73.4–79.9) | 12.9 (10.8–15.2) | 99.4 (98.3–99.7) | |
Original set + CycleGAN | 0.863 (0.811–0.914) | 80.7 (60.6–93.4) | 79.8 (76.5–82.8) | 13.4 (10.8–16.5) | 99.0 (97.9–99.5) | |
Original set + StyleGAN2 | 0.939 (0.899–0.979) | 84.6 (65.1–95.6) | 88.3 (85.6–90.6) | 22.0 (17.8–26.8) | 99.3 (98.3–99.7) | |
EfficientNetB0 | Original set (no augmentation) | 0.823 (0.759–0.886) | 96.1 (80.3–99.9) | 60.0 (56.2–63.8) | 8.5 (7.6–9.5) | 99.7 (98.3–99.9) |
Original set + classic augmentationa | 0.829 (0.799–0.857) | 80.8 (60.6–93.4) | 70.4 (66.8–73.8) | 9.6 (7.8–11.7) | 98.9 (97.7–99.5) | |
Original set + DCGAN | 0.875 (0.815–0.935) | 96.1 (80.3–99.9) | 64.1 (60.3–67.7_ | 9.4 (8.4–10.5) | 99.7 (98.4–99.9) | |
Original set + CycleGAN | 0.841 (0.773–0.908) | 100 (86.7–100) | 53.5 (49.6–57.3) | 7.7 (7.1–8.3) | 100 (98.9–100) | |
Original set + StyleGAN2 | 0.951 (0.926–0.976) | 96.1 (80.3–99.9) | 84.6 (81.6–87.2) | 19.5 (16.6–22.7) | 99.8 (98.8–99.9) |