Skip to main content

Table 2 Performance of segmentation 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 Model Dice Accuracy Patient-level sensitivity Specificity PPV NPV Lesion-level sensitivity FPR
Internal validation cohort Baseline 0.71 (0.68–0.74) 96.8 (96.4–97.2) 80.2 (77.6–82.7) 98.2 (97.8–98.5) 75.5 (72.3–78.5) 98.1 (97.8–98.4) 84.8 (81.9–87.8) 34.0 (30.5–37.4)
Baseline + TDL 0.72 (0.70–0.75) 96.7 (96.3–97.0) 83.3 (81.0–85.6) 97.7 (97.3–98.0) 73.4 (70.2–76.3) 98.4 (98.2–98.7) 81.8 (78.6–85.0) 32.2 (28.8–35.5)
Baseline + LRS 0.73 (0.70–0.76) 97.0 (96.6–97.4) 80.0 (77.3–82.8) 98.5 (98.2–98.7) 75.5 (72.3–78.4) 98.0 (97.7–98.4) 83.2 (80.1–86.3) 22.7 (19.3–26.1)
FFS 0.75 (0.73–0.78) 97.1 (96.8–97.5) 82.3 (79.8–84.8) 98.4 (98.1–98.6) 77.9 (75.0–80.6) 98.3 (97.9–98.6) 87.2 (84.4–89.9) 23.8 (20.4–27.3)
External validation cohort Baseline 0.71 (0.68–0.73) 96.8 (96.5–97.2) 73.1 (70.5–75.5) 99.3 (99.2–99.4) 83.3 (81.3–85.5) 97.2 (96.8–97.6) 90.1 (87.7–92.5) 34.7 (31.7–37.7)
Baseline + TDL 0.72 (0.70–0.75) 96.6 (96.3–97.0) 86.6 (84.8–88.4) 97.9 (97.6–98.1) 71.0 (68.5–73.7) 98.5 (98.1–98.8) 92.0 (89.8–94.2) 44.0 (40.9–47.0)
Baseline + LRS 0.71 (0.69–0.74) 96.9 (96.6–97.2) 78.2 (75.6–80.4) 98.7 (98.5–98.9) 77.0 (74.2–79.6) 97.9 (97.6–98.2) 92.5 (90.4–94.7) 34.1 (31.0–37.3)
FFS 0.73 (0.71–0.75) 97.1 (96.7–97.4) 79.0 (76.4–81.1) 98.7 (98.5–98.9) 79.6 (76.9–82.0) 98.1 (97.8–98.4) 94.3 (92.4–96.2) 30.8 (27.9–33.7)
  1. The data in parentheses are 95% confidence interval
  2. TDL Temporal difference learning, LRS Liver region segmentation, FFS Final fusion segmentation, PPV Positive predictive value, NPV Negative predictive value, FPR False-positives ratio