Skip to main content

Table 4 Quantitative evaluation of the combinations of cost-sensitive and data-level methods using CNN features

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

Methods ACC (%) SPC (%) SEN (%) F1_M (%) G_M (%) AUC (%)
ResCNN 90.22 (0.88)a 95.80 (1.23) 76.05 (3.21) 81.41 (1.74) 85.34 (1.59) 96.26 (0.73)
ResCNN + SMOTE 90.98 (1.07) 94.72 (1.34) 80.95 (3.50) 82.97 (2.05) 87.54 (1.75) 96.24 (0.84)
ResCNN + BSMOTE 90.76 (1.40) 95.48 (1.54) 78.10 (2.94) 82.12 (2.55) 86.34 (1.79) 96.27 (0.87)
ResCNN + UNDER 90.02 (1.68) 90.91 (1.71) 87.62 (3.71) 82.67 (2.82) 89.23 (2.14) 96.27 (0.80)
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)
CS-ResCNN + SMOTE 92.35 (1.01) 95.08 (0.93) 85.03 (4.91) 85.74 (2.21) 89.88 (2.31) 97.36 (0.70)
CS-ResCNN + BSMOTE 92.01 (0.86) 95.48 (1.04) 82.72 (3.78) 84.89 (1.83) 88.85 (1.80) 97.22 (0.65)
CS-ResCNN + UNDER 91.83 (0.85) 92.79 (1.74) 89.25 (3.80) 85.58 (1.42) 90.97 (1.41) 97.35 (0.61)
  1. ResCNN, residual convolutional neural network; CS-ResCNN, cost-sensitive residual convolutional neural network; SMOTE, synthetic minority over-sampling technique; BSMOTE, borderline-SMOTE; UNDER, under-sampling; CS-ResCNN + SMOTE, the combination of CS-ResCNN and SMOTE methods; CS-ResCNN + BSMOTE, the combination of CS-ResCNN and BSMOTE methods; CS-ResCNN + BSMOTE, the combination of CS-ResCNN and UNDER methods
  2. Italic represent the best value in all methods
  3. aMean (standard deviation)