Table 3.
Performance of the NDLS in the internal and external test datasets
| One-versus-rest classification | Internal test dataset |
External test dataset |
||
|---|---|---|---|---|
| High-quality images | Low-quality images | High-quality images | Low-quality images | |
| Keratitis versus others + normal | ||||
| Sensitivity (95% CI) | 95.4% (93.6–97.3) | 74.1% (64.5–83.6) | 90.7% (89.6–91.9) | 77.1% (69.1–85.2) |
| Specificity (95% CI) | 98.0% (96.8–99.2) | 92.9% (88.7–97.1) | 97.8% (97.3–98.3) | 90.9% (86.8–94.9) |
| Accuracy (95% CI) | 96.8% (95.6–97.9) | 86.0% (81.5–90.6) | 94.9% (94.3–95.4) | 86.1% (82.2–90.0) |
| Others versus keratitis + normal | ||||
| Sensitivity (95% CI) | 94.6% (90.7–98.5) | 84.1% (75.4–92.7) | 90.7% (89.2–92.2) | 80.5% (72.1–88.8) |
| Specificity (95% CI) | 97.4% (96.4–98.5) | 83.0% (77.1–89.0) | 93.6% (92.9–94.4) | 87.4% (83.0–91.9) |
| Accuracy (95% CI) | 97.1% (96.0–98.1) | 83.3% (78.4–88.2) | 92.9% (92.3–93.6) | 85.4% (81.5–89.4) |
| Normal versus keratitis + others | ||||
| Sensitivity (95% CI) | 99.2% (98.3–100) | 77.8% (68.2–87.4) | 94.0% (92.9–95.0) | 74.5% (66.4–82.7) |
| Specificity (95% CI) | 100% (100-100) | 92.0% (87.7–96.3) | 96.9% (96.3–97.4) | 87.5% (82.8–92.2) |
| Accuracy (95% CI) | 99.7% (99.4–100) | 87.4% (83.0–91.8) | 95.9% (95.4–96.4) | 82.8% (78.5–87.0) |
“Others” denotes other corneal abnormalities. “Normal” denotes normal cornea. NDLS, new deep learning system; CI, confidence interval.