Skip to main content

Table 3 Performance of predictive models in the validation cohorts

From: Real-time automatic prediction of treatment response to transcatheter arterial chemoembolization in patients with hepatocellular carcinoma using deep learning based on digital subtraction angiography videos

Cohort Input Model AUC Accuracy (%) Sensitivity (%) Specificity (%) PPV (%)  NPV (%)
Internal validation cohort KF Resnet 0.681 (0.637–0.725) 70.4 (65.3–75.0) 47.3 (39.5–55.2) 88.9 (84.5–93.0) 77.2 (69.1–85.4) 67.9 (61.5–73.5)
Clinical data MLP 0.670 (0.623–0.719) 67.7 (62.9–72.6) 60.0 (52.5–67.3) 73.9 (67.6–79.7) 64.7 (56.7–71.8) 69.9 (63.9–75.7)
KF+KF*Pred+Pred Resnet 0.733 (0.687–0.779) 73.7 (69.4–78.8) 69.7 (62.8–76.8) 76.8 (70.5–82.2) 70.6 (63.5–77.4) 76.1 (70.1–81.9)
KF+KF*Pred
+Pred+clinical data
Resnet+MLP 0.782 (0.738–0.826) 78.2 (74.2–82.3) 77.6 (70.7–84.0) 78.7 (72.9–84.1) 74.4 (67.2–81.4) 81.5 (75.9–86.8)
KF+ KF* GT+GT Resnet 0.727 (0.684–0.770) 74.7 (70.2–78.8) 55.2 (47.4–63.2) 90.3 (86.3–94.1) 82.0 (75.0–89.0) 71.6 (65.9–77.2)
KF+KF*GT
+GT+clinical data
Resnet+MLP 0.802 (0.759–0.847) 80.4 (76.3–84.4) 78.8 (72.2–84.9) 81.6 (76.5–86.6) 77.4 (70.6–83.7) 82.8 (77.4–88.0)
External validation cohort KF Resnet 0.628 (0.573–0.684) 69.4 (64.5–74.6) 49.5 (38.6–59.4) 76.2 (71.2–81.5) 41.4 (32.1–50.4) 81.6 (77.0–86.0)
Clinical data MLP 0.593 (0.529–0.646) 63.1 (58.5–68.3) 51.6 (42.1–63.1) 67.0 (61.6–72.7) 34.8 (26.8– 42.8) 80.3 (75.0–86.1)
KF+KF*Pred+Pred Resnet 0.712 (0.672–0.753) 73.9 (69.6–78.5) 46.7 (39.2–54.2) 95.7 (92.8–98.1) 89.5 (83.0–96.0) 69.2 (63.8–74.5)
KF+KF*Pred
+Pred+clinical data
Resnet+MLP 0.670 (0.612–0.726) 75.1 (70.2–79.5) 50.5 (40.0–61.5) 83.5 (79.2–87.7) 51.1 (40.6–61.4) 83.2 (78.9–87.5)
KF+KF*GT+GT Resnet 0.575 (0.536–0.614) 76.2 (71.6–80.6) 19.4 (12.1–28.3) 95.6 (93.0–97.9) 60.0 (41.2–78.3) 77.7 (73.3–81.9)
KF+KF*GT
+GT+clinical data
Resnet+MLP 0.817 (0.777–0.856) 77.9 (73.8–82.0) 89.2 (82.2–95.2) 74.0 (68.1–79.2) 53.9 (46.2–61.8) 95.3 (92.4–97.7)
  1. The data in parentheses are 95% confidence interval
  2. AUC Area under the curve, PPV Positive predictive value, NPV Negative predictive value, KF Key frame, Pred: segmentation result from Model 1; GT: segmentation result from ground truth; MLP Multi–layer perceptron