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

Ferreira, Mafalda Falcão; Camacho, Rui; Teixeira, Luís F.

doi:10.1186/s12911-020-01150-w

BMC Medical Informatics and Decision Making

Table 4 Performance comparison when using each of the 3 AEs — Basic AE, Denoising AE and Sparse AE — and for each type of cancer (Continued)

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

Lung	AE: Encoding Layers	85.97 ±7.00	0.54 ±0.13	65.00 ±17.54	61.25 ±12.30	60.94 ±11.01
	AE: Complete Autoencoder	90.93 ±2.56	0.67 ±0.09	77.28 ±9.43	66.90 ±8.26	71.51 ±7.94
	DAE: Encoding Layers	81.77 ±3.17	0.25 ±0.13	45.70 ±25.30	28.38 ±16.21	32.15 ±15.21
	DAE: Complete Autoencoder	85.73 ±3.28	0.49 ±0.09	60.30 ±9.76	53.40 ±7.49	56.21 ±7.44
	SAE: Encoding Layers	79.70 ±3.66	0.11 ±0.08	23.94 ±30.04	4.88 ±3.81	7.13 ±5.27
	SAE: Complete Autoencoder	83.23 ±2.59	0.40 ±0.09	51.33 ±7.62	49.33 ±10.08	49.83 ±7.52
Fine-Tuning the AE Weights (Approach B)
Thyroid	AE: Encoding Layers	99.67 ±0.42	0.99 ±0.01	98.29 ±2.09	99.80 ±0.62	99.03 ±1.21
	AE: Complete Autoencoder	99.67 ±0.22	0.99 ±0.01	99.22 ±1.00	98.82 ±1.02	99.02 ±0.65
	DAE: Encoding Layers	99.57 ±0.55	0.99 ±0.02	97.77 ±3.08	99.80 ±0.62	98.75 ±1.56
	DAE: Complete Autoencoder	99.60 ±0.38	0.99 ±0.01	99.22 ±1.01	98.42 ±2.05	98.81 ±1.15
	SAE: Encoding Layers	95.47 ±1.01	0.85 ±0.02	80.98 ±4.76	96.47 ±3.31	87.90 ±2.20
	SAE: Complete Autoencoder	97.73 ±0.52	0.93 ±0.02	89.39 ±2.69	98.43 ±2.03	93.65 ±1.41
Skin	AE: Encoding Layers	99.50 ±0.32	0.98 ±0.01	98.12 ±1.52	98.73 ±1.48	98.45 ±1.01
	AE: Complete Autoencoder	99.33 ±0.57	0.97 ±0.02	99.35 ±1.45	96.41 ±2.99	97.84 ±1.84
	DAE: Encoding Layers	99.30 ±0.51	0.97 ±0.02	97.52 ±2.12	98.09 ±2.34	97.78 ±1.62
	DAE: Complete Autoencoder	99.50 ±0.53	0.98 ±0.02	99.58 ±0.89	97.24 ±3.48	98.36 ±1.77
	SAE: Encoding Layers	95.80 ±1.18	0.84 ±0.05	93.23 ±5.06	79.43 ±7.22	85.51 ±4.38
	SAE: Complete Autoencoder	97.53 ±1.08	0.90 ±0.05	95.76 ±2.83	88.37 ±7.12	91.74 ±3.94
Stomach	AE: Encoding Layers	99.43 ±0.39	0.98 ±1.71	98.21 ±1.70	97.83 ±1.36	97.98 ±1.36
	AE: Complete Autoencoder	99.17 ±0.59	0.97 ±0.02	97.60 ±1.98	96.39 ±4.24	96.93 ±2.26
	DAE: Encoding Layers	99.33 ±0.47	0.97 ±0.02	97.84 ±2.10	97.35 ±2.39	97.57 ±1.72
	DAE: Complete Autoencoder	99.23 ±0.57	0.97 ±0.02	98.08 ±1.90	96.35 ±3.86	97.16 ±2.16
	SAE: Encoding Layers	95.60 ±0.81	0.81 ±0.04	93.33 ±3.92	73.72 ±7.08	82.12 ±3.96
	SAE: Complete Autoencoder	97.37 ±0.55	0.89 ±2.89	96.08 ±3.01	84.56 ±4.90	89.83 ±2.43
Breast	AE: Encoding Layers	99.33 ±0.52	0.98 ±0.02	97.85 ±2.32	98.20 ±1.48	98.01 ±1.55
	AE: Complete Autoencoder	99.30 ±0.37	0.98 ±0.01	99.00 ±1.06	96.80 ±2.35	97.87 ±1.15
	DAE: Encoding Layers	99.20 ±0.65	0.97 ±0.02	97.83 ±2.54	97.40 ±1.90	97.60 ±1.95
	DAE: Complete Autoencoder	99.23 ±0.52	0.97 ±0.02	98.60 ±2.08	96.80 ±1.69	97.68 ±1.57
	SAE: Encoding Layers	96.70 ±1.24	0.89 ±0.05	95.29 ±4.78	84.60 ±6.47	89.45 ±4.14
	SAE: Complete Autoencoder	97.40 ±1.12	0.90 ±0.04	95.78 ±4.02	88.40 ±4.79	91.87 ±3.52
Lung	AE: Encoding Layers	99.27 ±0.83	0.97 ±0.03	97.34 ±3.08	98.44 ±2.02	97.87 ±2.40
	AE: Complete Autoencoder	99.23 ±0.45	0.97 ±0.02	98.83 ±1.63	96.67 ±2.46	97.71 ±1.34
	DAE: Encoding Layers	99.00 ±0.75	0.96 ±0.03	96.89 ±2.27	97.26 ±2.65	97.06 ±2.23
	DAE: Complete Autoencoder	99.27 ±0.52	0.97 ±0.02	97.95 ±2.69	97.85 ±3.12	97.87 ±1.58
	SAE: Encoding Layers	95.27 ±1.43	0.82 ±0.06	90.69 ±4.64	80.61 ±6.72	85.21 ±4.78
	SAE: Complete Autoencoder	97.00 ±0.96	0.89 ±0.04	93.65 ±2.56	88.44 ±5.36	90.88 ±3.19

When measuring loss, lower is better. For all the remaining metrics, higher is better. 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 highlighted values correspond to the combination that led to the overall best result (detecting thyroid cancer, importing only the encoding layers a Basic AE into the classification network, and allowing subsequent fine-tune, when training for the classification task)

Back to article page

ISSN: 1472-6947

Contact us

General enquiries: journalsubmissions@springernature.com