Skip to main content

Table 5 Comparison of final models across all five folds for binary models

From: A radiographic, deep transfer learning framework, adapted to estimate lung opacities from chest x-rays

Model Architecture

Data Balancing Strategy

Segmentation Strategy

Precision Mean (Stdev)

Recall Mean (Stdev)

F1 Mean (Stdev)

(A) None versus mild/medium/severe

 CheXNet-121

DOUBLE

NONE

93.47 (0.92)

88.07 (0.78)

90.69 (0.7)

 CheXNet-121

UNDER

NONE

92.70 (0.93)

88.76 (1.29)

90.68 (0.78)

 ResNet-50

UNDER

NONE

92.61 (1.7)

88.08 (1.86)

90.25 (0.68)

 ResNet-50

DOUBLE

SPINE

92.05 (2.03)

88.18 (2.15)

90.03 (0.63)

 VGG-16

DOUBLE

SPINE

79.63 (1.85)

79.63 (2.29)

81.46 (0.82)

(B) None/mild versus medium/severe

 CheXNet-121

DOUBLE

NONE

80.74 (1.38)

81.91 (2.51)

81.29 (1.44)

 CheXNet-121

UNDER

NONE

78.70 (1.83)

84.81 (1.6)

81.62 (1.1)

 ResNet-50

UNDER

NONE

80.06 (2.42)

81.24 (4.72)

80.46 (1.66)

 ResNet-50

DOUBLE

SPINE

79.12 (1.7)

81.65 (4.09)

80.24 (1.41)

 VGG-16

DOUBLE

SPINE

79.63 (2.14)

83.38 (4.27)

79.64 (2.07)

(C) None/mild/medium versus severe

 CheXNet-121

DOUBLE

NONE

62.91 (3.34)

59.32 (4.83)

60.91 (2.91)

 CheXNet-121

UNDER

NONE

58.96 (4.09)

64.97 (3.24)

61.61 (2.27)

 ResNet-50

UNDER

NONE

64.43 (3.8)

54.61 (8.26)

58.42 (4.16)

 ResNet-50

DOUBLE

SPINE

62.05 (3.81)

57.43 (6.92)

59.24 (3.85)