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) | 0.766 (0.702–0.830) | 71.4 (58.6–82.1) | 70.4 (64.3–75.9) | 37.8 (32.2–43.7) | 90.7 (86.7–93.6) |
Original set + classic augmentationa | 0.827 (0.780–0.867) | 92.1 (82.4–97.4) | 58.0 (51.6–64.2) | 35.6 (31.9–39.4) | 96.7 (92.5–98.5) | |
Original set + DCGAN | 0.850 (0.798–0.902) | 90.4 (80.4–96.4) | 63.6 (57.3–69.5) | 38.5 (34.2–42.9) | 96.3 (92.4–98.2) | |
Original set + CycleGAN | 0.859 (0.803–0.914) | 79.3 (67.3–88.5) | 78.8 (73.2–83.6) | 48.5 (41.8–55.2) | 93.8 (90.2–96.1) | |
Original set + StyleGAN2 | 0.913 (0.872–0.954) | 90.4 (80.4–96.4) | 75.6 (69.7–80.7) | 48.3 (42.5–54.1) | 96.9 (93.6–98.5) | |
EfficientNetB0 | Original set (no augmentation) | 0.796 (0.736–0.855) | 71.4 (58.6–82.1) | 71.6 (65.5–77.1) | 38.7 (33.0–44.8) | 90.8 (86.9–93.6) |
Original set + classic augmentationa | 0.833 (0.786–0.872) | 88.9 (78.4–95.4) | 62.4 (56.1–68.4) | 37.3 (33.2–41.6) | 95.7 (91.7–97.8) | |
Original set + DCGAN | 0.821 (0.756–0.887) | 65.0 (52.0–76.6) | 88.4 (83.7–92.0) | 58.5 (48.9–67.5) | 90.9 (87.7–93.3) | |
Original set + CycleGAN | 0.875 (0.821–0.929) | 71.4 (58.6–82.1) | 90.4 (86.0–93.7) | 65.2 (55.4–73.8) | 92.6 (89.4–94.8) | |
Original set + StyleGAN2 | 0.926 (0.890–0.963) | 92.0 (82.4–97.3) | 80.8 (75.3–85.4) | 54.7 (48.1–61.1) | 97.5 (94.5–98.9) |