Table 3 Diagnostic performance of the convolutional neural network computer-aided detection system compared to bronchoscopist.
From: Deep learning-based diagnosis from endobronchial ultrasonography images of pulmonary lesions
|  | CNN-CAD | CNN-CAD | Bronchoscopist | Total (N = 4) | |||
|---|---|---|---|---|---|---|---|
(All images) | (Each case) | Expert 1 | Expert 2 | Trainee 1 | Trainee 2 | ||
Accuracy, % (95% CI) | 83.4 (83.0–83.9) | 83.3 (68.6–93.0) | 73.8 (58.0–86.1) | 66.7 (50.5–80.4) | 57.1 (41.0–72.3) | 76.2 (60.5–87.9) | 68.5 (60.8–75.4) |
Sensitivity, % (95% CI) | 95.3 (95.0–95.6) | 100 (83.3–100) | 80 (61.4–92.3) | 83.3 (65.3–94.4) | 63.3 (43.9–80.1) | 93.3 (77.9–99.2) | 80.0 (71.7–86.7) |
Specificity, % (95% CI) | 53.4 (52.3–54.6) | 41.7 (15.2–72.3) | 58.3 (27.7–84.8) | 25 (5.5–57.2) | 41.7 (15.2–72.3) | 33.3 (9.9–65.1) | 39.6 (25.8–54.7) |
PPV, % (95% CI) | 83.8 (83.3–84.3) | 81.1 (64.8–92.0) | 82.8 (64.2–94.2) | 73.5 (55.6–87.1) | 73.1 (52.2–88.4) | 77.8 (60.8–89.9) | 76.8 (68.4–83.9) |
NPV, % (95% CI) | 82.0 (80.9–83.0) | 100 (35.9–100) | 53.8 (25.1–80.8) | 37.5 (8.5–75.5) | 31.2 (11.0–58.7) | 66.7 (22.3–95.7) | 44.2 (29.1–60.1) |
