Table 9.

Evaluation results of 110 testing cases in BRATS 2015 testing dataset (the sizes of image patches used to train FCNNs were 33^*33^*3 and 65^*65^*3 respectively, n = 5). Models 2013: models trained based on the BRATS 2013 training dataset; Models 2015: models trained based on the BRATS 2015 training dataset.

	Methods		Dice			PPV			Sensitivity
			Complete	Core	Enhancing	Complete	Core	Enhancing	Complete	Core	Enhancing
Models 2013	FCNN + CRF	axial	0.77	0.56	0.52	0.85	0.71	0.46	0.74	0.53	0.67
		coronal	0.74	0.53	0.48	0.70	0.69	0.44	0.84	0.50	0.65
		sagittal	0.74	0.54	0.49	0.74	0.67	0.43	0.78	0.52	0.66
	fusing (FCNN + CRF)		0.79	0.56	0.56	0.82	0.78	0.52	0.79	0.50	0.68
	FCNN + CRF + post-process	axial	0.78	0.64	0.58	0.87	0.75	0.57	0.73	0.61	0.65
		coronal	0.78	0.61	0.57	0.77	0.68	0.58	0.84	0.63	0.63
		sagittal	0.77	0.62	0.57	0.80	0.71	0.56	0.78	0.60	0.63
	fusing (FCNN + CRF) + post-process		0.81	0.65	0.60	0.87	0.79	0.60	0.78	0.61	0.65
Models 2015	FCNN + CRF	axial	0.78	0.64	0.54	0.78	0.76	0.48	0.81	0.62	0.71
		coronal	0.77	0.66	0.56	0.73	0.73	0.52	0.86	0.67	0.67
		sagittal	0.76	0.63	0.47	0.75	0.71	0.38	0.80	0.63	0.75
	fusing (FCNN + CRF)		0.80	0.66	0.57	0.81	0.79	0.50	0.83	0.64	0.72
	FCNN + CRF + post-process	axial	0.80	0.68	0.61	0.82	0.75	0.59	0.81	0.71	0.68
		coronal	0.80	0.69	0.61	0.77	0.70	0.61	0.87	0.76	0.65
		sagittal	0.78	0.69	0.57	0.80	0.74	0.51	0.80	0.71	0.72
	fusing (FCNN + CRF) + post-process		0.82	0.72	0.62	0.84	0.78	0.60	0.83	0.73	0.69