Skip to main content

Table 3 Quantitative evaluation of the CS-ResCNN method and various conventional methods

From: Automatic diagnosis of imbalanced ophthalmic images using a cost-sensitive deep convolutional neural network

Methods

ACC (%)

SPC (%)

SEN (%)

F1_M (%)

G_M (%)

AUC (%)

WT

75.49 (0.70)a

99.29 (0.90)

11.70 (1.55)

20.57 (2.42)

34.02 (2.26)

82.50 (1.93)

WT-SMOTE

78.34 (1.35)

89.34 (0.97)

48.84 (2.52)

55.06 (2.81)

66.05 (2.04)

83.18 (2.05)

WT-BSMOTE

78.85 (1.48)

91.73 (0.79)

44.35 (3.48)

53.24 (3.72)

63.75 (2.79)

83.89 (2.37)

WT-UNDER

72.98 (2.07)

70.00 (2.63)

80.95 (3.37)

61.96 (2.36)

75.25 (2.02)

83.40 (2.38)

LBP

81.70 (0.68)

94.67 (0.78)

46.94 (3.15)

58.18 (2.35)

66.63 (2.06)

86.04 (1.66)

LBP-SMOTE

82.40 (1.37)

91.52 (1.37)

57.96 (2.38)

64.16 (2.58)

72.82 (1.76)

86.78 (1.81)

LBP-BSMOTE

82.18 (1.07)

92.18 (1.13)

55.37 (1.77)

62.81 (2.03)

71.44 (1.36)

86.57 (1.80)

LBP-UNDER

79.22 (2.16)

80.81 (1.82)

74.97 (4.45)

66.22 (3.50)

77.81 (2.80)

86.33 (1.92)

SIFT

83.81 (0.91)

96.24 (1.05)

50.48 (2.12)

62.88 (2.01)

69.69 (1.47)

91.14 (1.44)

SIFT-SMOTE

85.43 (2.20)

92.59 (1.24)

66.26 (4.94)

71.16 (4.56)

78.30 (3.37)

91.34 (1.32)

SIFT-BSMOTE

84.88 (1.21)

92.74 (1.44)

63.81 (2.86)

69.63 (2.32)

76.91 (1.74)

91.31 (1.34)

SIFT-UNDER

81.29 (1.57)

79.54 (1.85)

85.99 (2.07)

71.43 (2.02)

82.69 (1.52)

91.29 (1.11)

COTE

78.15 (1.10)

92.34 (1.26)

40.14 (3.97)

49.89 (3.47)

60.81 (2.91)

82.19 (2.34)

COTE-SMOTE

76.60 (1.07)

80.46 (0.74)

66.26 (3.35)

60.59 (2.26)

72.99 (1.92)

81.41 (1.68)

COTE-BSMOTE

76.75 (1.32)

80.71 (0.65)

66.12 (3.83)

60.68 (2.72)

73.03 (2.28)

80.98 (1.61)

COTE-UNDER

72.09 (1.52)

69.85 (1.65)

78.10 (3.74)

60.32 (2.24)

73.83 (1.93)

80.87 (1.73)

ResCNN

90.22 (0.88)

95.80 (1.23)

76.05 (3.21)

81.41 (1.74)

85.34 (1.59)

96.26 (0.73)

CS-ResCNN

92.24 (1.30)

93.19 (1.73)

89.66 (2.86)

86.00 (2.27)

91.39 (1.49)

97.11 (0.59)

  1. The random forest classifier is employed for the conventional methods. ResCNN, residual convolutional neural network; CS-ResCNN, cost-sensitive residual convolutional neural network; WT, wavelet transformation; LBP, local binary pattern; SIFT, scale-invariant feature transform; COTE, color and texture features; SMOTE, synthetic minority over-sampling technique; BSMOTE, borderline-SMOTE; UNDER, under-sampling; ACC, accuracy; SPC, specificity; SEN, sensitivity; F1_M, F1-measure; G_M, G-mean; AUC, area under the receiver operating characteristic curve
  2. Italics represents the best value in all methods
  3. aMean (standard deviation)