Table 2 Performance of previously established AI diagnostic systems in images of varying quality.
From: Deep learning from “passive feeding” to “selective eating” of real-world data
Zhongshan ophthalmic centre dataset | Xudong ophthalmic hospital dataset | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
GON | RED | LDRB | GON | RED | ||||||
Sensitivity (95% CI) | Specificity (95% CI) | Sensitivity (95% CI) | Specificity (95% CI) | Sensitivity (95% CI) | Specificity (95% CI) | Sensitivity (95% CI) | Specificity (95% CI) | Sensitivity (95% CI) | Specificity (95% CI) | |
Good and poor quality (without DLIFS) | 88.6% (85.1–92.1) | 95.7% (94.6–96.8) | 89.1% (85.5–92.7) | 94.0% (92.7–95.3) | 86.6% (81.9–91.3) | 94.6% (93.8–95.4) | 90.3% (87.3–93.3) | 97.9% (97.3–98.5) | 85.1% (80.8–89.4) | 97.3% (96.7–97.9) |
Good quality only (with DLIFS) | 98.1% (96.4–98.8) | 96.6% (95.5–97.7) | 94.9% (92.1–97.7) | 95.1% (93.8–96.4) | 96.2% (93.2–99.2) | 97.6% (97.0–98.2) | 98.0% (96.4–99.6) | 98.5% (98.0–99.0) | 95.8% (92.9–98.7) | 98.0% (97.4–98.6) |
Poor quality only | 52.2% (40.2–64.2) | 90.3% (85.9–94.7) | 67.2% (55.4–79.0) | 87.9% (83.1–92.7) | 53.3% (38.7–67.9) | 67.3% (61.7–72.9) | 58.1% (46.9–69.3) | 92.6% (89.3–95.9) | 59.0% (48.1–69.9) | 90.0% (86.2–93.8) |
