Table 2.
Per-class performances of the deep learning models in the four-class classification.
| Model/Class | Sensitivity (%) | Specificity (%) | PPV (%) | NPV (%) | F1 Score | AUC (95% CI) |
|---|---|---|---|---|---|---|
| DenseNet-161 | ||||||
| CIN3 | 95.3 (93.7–96.8) | 94.4 (93.1–95.6) | 85.0 (82.3–87.7) | 98.3 (97.8–98.9) | 89.8 (88.8–90.8) | 0.989 (0.982–0.996) |
| CIN2 | 75.2 (67.7–82.8) | 94.1 (91.8–96.4) | 76.1 (62.3–89.9) | 93.8 (92.5–95.0) | 75.5 (64.9–86.1) | 0.947 (0.932–0.963) |
| CIN1 | 82.1 (77.8–86.5) | 98.3 (97.4–99.2) | 94.2 (92.2–96.2) | 94.5 (92.6–96.4) | 87.7 (84.5–91.0) | 0.979 (0.968–0.990) |
| Non-neoplasm | 95.6 (90.9–100.0) | 98.0 (96.3–99.7) | 95.0 (91.0–99.0) | 98.3 (96.6–100.0) | 95.2 (92.0–98.4) | 0.996 (0.991–1.000) |
| EfficientNet-B7 | ||||||
| CIN3 | 97.5 (95.4–99.5) | 96.3 (94.1–98.6) | 90.0 (84.2–95.8) | 99.1 (98.4–99.8) | 93.6 (89.6–97.5) | 0.990 (0.981–0.999) |
| CIN2 | 73.0 (62.2–83.9) | 96.7 (93.7–99.7) | 86.8 (75.2–98.4) | 93.6 (92.3–94.8) | 79.1 (69.1–89.1) | 0.956 (0.946–0.967) |
| CIN1 | 85.2 (73.3–97.1) | 96.3 (95.1–97.6) | 88.5 (88.2–88.8) | 95.5 (91.3–99.8) | 86.5 (80.4–92.6) | 0.971 (0.950–0.993) |
| Non-neoplasm | 95.6 (90.9–100.0) | 96.3 (92.2–100.0) | 92.3 (84.8–99.8) | 98.3 (96.6–100.0) | 93.8 (88.8–98.8) | 0.996 (0.992–0.999) |
PPV, positive predictive value; NPV, negative predictive value; AUC, area under the receiver operating characteristic curve; CI, confidence interval.