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

Table 5 Performance comparison when using each of the 3 AEs — Basic AE, Denoising AE and Sparse AE — and for malaria detection

	Top Layers (AE)	Accuracy (%)	MCC	Precision (%)	Recall (%)	F₁ score
Approach A	AE: Encoding Layers	62.82 ±0.60	0.03 ±0.04	66.17 ±41.60	1.90 ±3.97	3.32 ±6.49
	AE: Complete Autoencoder	62.73 ±0.45	0.03 ±0.04	53.97 ±34.95	2.12 ±4.23	3.66 ±6.72
	DAE: Encoding Layers	62.50 ±0.07	0.00 ±0.00	0.00 ±0.00	0.00 ±0.00	0.00 ±0.00
	DAE: Complete Autoencoder	62.50 ±0.07	0.00 ±0.00	0.00 ±0.00	0.00 ±0.00	0.00 ±0.00
	SAE: Encoding Layers	62.21 ±0.34	-0.01 ±0.02	13.17 ±17.36	0.23 ±0.32	0.44 ±0.63
	SAE: Complete Autoencoder	62.51 ±0.07	0.21 ±0.65	10.00 ±31.62	0.05 ±0.14	0.09 ±0.28
Approach B	AE: Encoding Layers	91.28 ±1.17	0.82 ±0.02	87.41 ±2.69	89.84 ±3.32	88.53 ±1.58
	AE: Complete Autoencoder	91.43 ±1.21	0.82 ±0.02	88.12 ±2.37	89.25 ±2.21	88.66 ±1.58
	DAE: Encoding Layers	92.36 ±0.46	0.84 ±0.01	88.91 ±2.14	91.11 ±2.49	89.95 ±0.61
	DAE: Complete Autoencoder	92.18 ±0.83	0.84 ±0.02	88.54 ±2.33	91.19 ±1.68	89.78 ±1.00
	SAE: Encoding Layers	62.21 ±0.34	-0.01 ±0.02	13.17 ±17.36	0.23 ±0.32	0.44 ±0.63
	SAE: Complete Autoencoder	62.51 ±0.07	0.01 ±0.01	10.00 ±31.62	0.05 ±0.14	0.09 ±0.28

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 Denoising AE into the classification network, and allowing subsequent fine-tune, when training for the classification task)

ISSN: 1472-6947