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

Table 6 Performance comparison when using each of the 3 AEs — Basic AE, Denoising AE and Sparse AE — and for breast cancer detection, on the UCI’s Wisconsin Breast Cancer dataset

	Top Layers (AEs)	Accuracy (%)	MCC	Precision (%)	Recall (%)	F₁ score
Approach A	AE: Encoding Layers	97.54 ±2.06	0.95 ±0.04	98.67 ±2.92	94.81 ±5.20	96.60 ±2.90
	AE: Complete Autoencoder	96.49 ±2.62	0.93 ±0.05	96.83 ±3.71	93.83 ±7.12	95.11 ±3.93
	DAE: Encoding Layers	95.43 ±3.81	0.90 ±0.08	98.38 ±3.48	89.13 ±8.73	93.36 ±6.05
	DAE: Complete Autoencoder	93.32 ±3.78	0.86 ±0.08	98.19 ±2.93	83.46 ±8.52	90.09 ±5.98
	SAE: Encoding Layers	97.19 ±2.22	0.94 ±0.05	97.69 ±3.15	94.81 ±5.20	96.14 ±3.13
	SAE: Complete Autoencoder	97.02 ±2.35	0.94 ±0.05	97.70 ±2.42	94.31 ±7.03	95.80 ±3.64
Approach B	AE: Encoding Layers	99.12 ±1.24	0.98 ±0.03	98.71 ±2.86	99.05 ±2.01	98.84 ±1.59
	AE: Complete Autoencoder	98.60 ±1.38	0.97 ±0.03	97.75 ±2.38	98.57 ±3.21	98.11 ±1.91
	DAE: Encoding Layers	97.72 ±2.62	0.95 ±0.06	98.08 ±2.50	95.74 ±6.13	96.81 ±3.83
	DAE: Complete Autoencoder	97.19 ±2.64	0.94 ±0.06	96.39 ±4.57	96.23 ±4.91	96.22 ±3.62
	SAE: Encoding Layers	97.19 ±2.22	0.94 ±0.04	96.15 ±5.24	96.71 ±3.19	96.31 ±2.78
	SAE: Complete Autoencoder	96.66 ±2.10	0.93 ±0.04	97.66 ±3.28	93.44 ±5.47	95.39 ±2.96

All the presented results are the 10-fold cross-validation mean values, at the validation set, by selecting the best performing model according to its F₁ score. The first row presents the results for Approach A, where we fix the resulting weights of the AE pre-training; the second one shows the results for Approach B, where we allow the subsequent fine-tuning of all the weights of the model. The highlighted values correspond to the combination that led to the overall best result (importing only the encoding layers a Basic AE into the classification network, and allowing subsequent fine-tune, when training for the classification task.)

ISSN: 1472-6947