A Combined Fully Convolutional Networks and Deformable Model for Automatic Left Ventricle Segmentation Based on 3D Echocardiography

Table 3

The comparison results of the proposed methods and the five state-of-the-art segmentation methods (Barbosa et al. [28], Milletari et al. [30], van Stralen et al. [31], Smistad et al. [40], and Keraudren et al. [41]) on 3DE in test set.

Method	ED						ES
					dice						dice
	mean	std	mean	std	mean	std	mean	std	mean	std	mean	std
[28]	2.26	0.73	8.10	2.66	0.106	0.041	2.43	0.91	8.13	3.08	0.144	0.057
[30]	2.14	0.68	8.25	3.87	0.107	0.031	2.91	1.01	8.53	2.30	0.162	0.062
[31]	2.44	0.91	8.45	3.50	0.121	0.054	2.79	1.24	8.65	2.85	0.165	0.079
[40]	2.62	0.95	8.26	2.98	0.115	0.038	2.92	0.93	8.99	2.98	0.156	0.050
[41]	2.44	0.95	8.98	3.09	0.130	0.048	2.54	0.75	9.15	3.24	0.158	0.057
Standard initialization + improved snake	4.7	1.76	13.2	4.9	0.22	0.15	4.9	1.7	11.2	5.7	0.25	0.16
improved FCN	3.1	0.8	8.7	3.2	0.17	0.09	2.9	0.57	9.56	3.6	0.21	0.07
improved FCN+ level set	3.2	0.6	9.1	2.9	0.11	0.07	2.5	0.57	9.9	3.7	0.17	0.04
improved FCN+ improved snake	2.03	0.41	8.80	3.69	0.098	0.0007	2.35	0.63	9.09	3.42	0.125	0.0008