Table 3 Performance comparisons of the proposed EMV-3D-CNN model and the six radiologists on the validation dataset for Task 3.

From: An ensemble deep learning model for risk stratification of invasive lung adenocarcinoma using thin-slice CT

Overall accuracy (%)

Task 3

Model

D1

D2

D3

D4

D5

D6

77.6

67.1

69.7

61.8

59.2

56.6

52.6

G1

G2

G3

G1

G2

G3

G1

G2

G3

G1

G2

G3

G1

G2

G3

G1

G2

G3

G1

G2

G3

Accuracy (%)

88.2

78.9

88.2

78.9

72.4

82.9

73.7

78.9

86.8

75.0

71.1

77.6

72.4

68.4

77.6

71.1

63.2

78.9

73.7

53.9

77.6

Sensitivity (%)

93.1

69.2

66.7

89.7

38.5

71.4

93.1

50.0

61.9

79.3

42.3

61.9

86.2

34.6

52.4

75.9

30.8

61.9

34.5

42.3

90.5

Specificity (%)

85.1

84.0

96.4

72.3

90.0

87.3

61.7

94.0

96.4

72.3

86.0

83.6

63.8

86.0

87.3

68.1

80.0

85.5

97.9

60.0

72.7

PPV (%)

79.4

69.2

87.5

66.7

66.7

68.2

60.0

81.3

86.7

63.9

61.1

59.1

59.5

56.3

61.1

59.5

44.4

61.9

90.9

35.5

55.9

NPV (%)

95.2

84.0

88.3

91.9

91.9

88.9

93.5

78.3

86.9

85.0

74.1

85.2

88.2

71.7

82.8

82.1

69.0

85.5

70.8

66.7

95.2

F1 (%)

85.7

69.2

75.7

76.5

76.5

69.8

73.0

61.9

72.2

70.1

50.0

60.5

70.4

42.9

56.4

66.7

36.4

61.9

50.0

38.6

69.1

  1. PPV positive predictive value, NPV negative predictive value, D Doctor, G Grade.