Table 3 The prediction results from the external validation dataset (JSIEC) 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) | 0.808 (0.703–0.912) | 88.4 (69.8–97.5) | 60.5 (43.3–75.9) | 60.5 (50.2–69.9) | 88.4 (71.9–95.8) |
Original set + classic augmentationa | 0.849 (0.738–0.926) | 92.3 (74.9–99.1) | 71.1 (54.1–84.6) | 68.6 (56.7–78.4) | 93.1 (77.8–98.1) | |
Original set + DCGAN | 0.866 (0.753–0.980) | 92.3 (74.8–99.0) | 76.3 (59.7–88.5) | 72.7 (59.8–82.6) | 93.5 (79.0–98.2) | |
Original set + CycleGAN | 0.879 (0.796–0.961) | 69.2 (48.2–85.6) | 89.4 (75.1–97.0) | 81.8 (63.2–92.1) | 80.9 (70.2–88.4) | |
Original set + StyleGAN2 | 0.910 (0.817–0.983) | 76.9 (56.3–91.0) | 92.1 (78.6–98.3) | 86.9 (68.8–95.2) | 85.3 (74.1–92.2) | |
EfficientNetB0 | Original set (no augmentation) | 0.890 (0.794–0.985) | 84.6 (65.1–95.6) | 92.1 (78.6–98.3) | 88.0 (70.9–95.6) | 89.7 (77.9–95.5) |
Original set + classic augmentationa | 0.899 (0.798–0.960) | 84.6 (65.1–95.6) | 94.7 (82.3–99.4) | 91.7 (73.9–97.7) | 90.0 (78.5–95.7) | |
Original set + DCGAN | 0.861 (0.764–0.959) | 92.3 (74.8–99.0) | 76.3 (59.7–88.5) | 72.7 (59.8–82.6) | 93.5 (79.0–98.2) | |
Original set + CycleGAN | 0.871 (0.778–0.967) | 76.9 (56.3–91.0) | 86.8 (71.9–95.5) | 80.0 (63.2–90.2) | 84.6 (72.9–91.8) | |
Original set + StyleGAN2 | 0.914 (0.818–0.999) | 88.4 (69.8–97.5) | 94.7 (82.2–99.3) | 92.0 (74.7–97.8) | 92.3 (80.5–97.2) |