Table 3 Performance of PAN-VIQ in vessel-specific vascular invasion prediction: Internal and external retrospective validation across five target vessels

From: A clinically validated 3D deep learning approach for quantifying vascular invasion in pancreatic cancer

Vessel

Cohort

Classification

Accuracy

Precision

Recall

F1 Score

Specificity

CA

Ruijin Hospital

0

97.54%

99.59%

98.78%

99.18%

97.03%

1

87.22%

91.81%

89.46%

98.71%

2

74.29%

80.00%

77.04%

99.05%

External validation cohort

0

95.93%

98.09%

98.09%

98.09%

80.00%

1

71.43%

71.43%

71.43%

98.79%

2

75.00%

75.00%

75.00%

98.78%

CHA

Ruijin Hospital

0

95.87%

99.64%

98.31%

98.97%

97.58%

1

76.19%

81.75%

78.87%

98.08%

2

67.74%

75.68%

71.49%

97.84%

External validation cohort

0

92.90%

97.33%

98.65%

97.99%

80.95%

1

55.56%

45.45%

50.00%

97.47%

2

60.00%

60.00%

60.00%

97.48%

SMA

Ruijin Hospital

0

94.38%

96.67%

98.78%

97.71%

86.78%

1

86.26%

83.09%

84.64%

97.87%

2

80.00%

65.12%

71.79%

98.85%

External validation cohort

0

97.01%

98.08%

99.35%

98.71%

76.92%

1

75.00%

75.00%

75.00%

98.74%

2

100.00%

60.00%

75.00%

100.00%

SMV

Ruijin Hospital

0

93.30%

96.63%

97.18%

96.90%

88.46%

1

83.08%

76.21%

79.50%

97.46%

2

79.79%

86.52%

83.02%

97.93%

External validation cohort

0

94.67%

98.50%

97.76%

98.13%

94.29%

1

84.21%

88.89%

86.49%

98.01%

2

76.47%

76.47%

76.47%

97.37%

PV

Ruijin Hospital

0

90.82%

94.76%

97.09%

95.91%

80.00%

1

75.37%

66.23%

70.51%

97.11%

2

72.99%

69.02%

70.95%

97.36%

External validation cohort

0

93.49%

96.38%

97.79%

97.08%

84.85%

1

75.00%

54.55%

63.16%

98.73%

2

82.61%

96.36%

84.44%

97.28%

  1. The PAN-VIQ model demonstrates robust performance across all vessel classifications (0: no invasion; 1: ≤180° encasement; 2: >180° encasement).
  2. CA celiac artery, CHA common hepatic artery, SMA superior mesenteric artery, SMV superior mesenteric vein, PV portal vein.