Table 2.
Performance of four deep learning algorithms in identifying malignant eyelid tumors in the internal and external test sets.
| Deep learning algorithms | Internal test set | External test set | ||||
|---|---|---|---|---|---|---|
| Sensitivity (95% CI) | Specificity (95% CI) | Accuracy (95% CI) | Sensitivity (95% CI) | Specificity (95% CI) | Accuracy (95% CI) | |
| DenseNet121 | 96.1% (90.8–100) | 77.4% (70.6–84.2) | 82.2% (76.9–87.6) | 91.5% (84.4–98.6) | 79.2% (73.9–84.5) | 81.8% (77.3–86.2) |
| ResNet50 | 94.1% (87.7–100) | 77.4% (70.6–84.2) | 81.7% (76.3–87.1) | 69.5% (57.7–81.2) | 84.5% (79.8–89.2) | 81.4% (76.9–85.9) |
| Inception-v3 | 90.2% (82.0–98.4) | 77.4% (70.6–84.2) | 80.7% (75.2–86.2) | 74.6% (63.5–85.7) | 73.9% (68.2–79.6) | 74.0% (68.9–79.1) |
| VGG16 | 86.3% (76.8–95.7) | 78.8% (72.1–85.4) | 80.7% (75.2–86.2) | 66.1% (54.0–78.2) | 82.3% (77.3–87.3) | 78.9% (74.2–83.7) |
CI, confidence interval.