Skip to main content

Table 8 Performance comparison when using the vanilla AE on two held-out test sets (Malaria and Wisconsin Breast Cancer, respectively)

From: Using autoencoders as a weight initialization method on deep neural networks for disease detection

 

Top Layers (AEs)

Accuracy (%)

MCC

Precision (%)

Recall (%)

F1 score

Train

M: Encoding Layers

91.21% ±1.56%

0.81 ±0.03

86.61% ±3.77%

90.84% ±2.91%

88.59% ±1.86%

 

M: Complete Autoencoder

90.19% ±2.08%

0.80 ±0.04

85.29% ±4.69%

89.69% ±2.83%

87.33% ±2.33%

 

WBC: Encoding Layers

98.69% ±1.38%

0.97 ±0.03

99.37% ±1.98%

97.14% ±3.69%

98.20% ±1.91%

 

WBC: Complete Autoencoder

97.90% ±2.07%

0.96 ±0.04

95.54% ±5.18%

99.29% ±2.26%

97.28% ±2.65%

Test

M: Encoding Layers

89.64%

0.78

90.02%

81.40%

85.49%

 

M: Complete Autoencoder

86.10%

0.70

82.86%

79.34%

81.06%

 

WBC: Encoding Layers

97.34%

0.95

99.99%

92.86%

96.30%

 

WBC: Complete Autoencoder

95.74%

0.91

98.44%

90.00%

94.03%

  1. The presented results in the first row (Train) are the 10-fold cross-validation mean values, at the validation set, by selecting the best performing model according to its F1 score. The second row (Test) gathers the results when evaluating the models on the testing phase. For both datasets, two thirds of the data were used in the training phase, and one third as the held-out in the test phase. M represents the Malaria dataset, and WBC the Wisconsin Breast Cancer one