Utility of deep learning networks for the generation of artificial cardiac magnetic resonance images in congenital heart disease

Table 1 Comparison between the segmentation accuracy

Cardiac Chamber	Pg-GAN	Actual pat. MRI	*p-value*
Percent Variation
Long axis view:
Left Ventricle	0.021 [0.017–0.027]	0.014 [0.012–0.018]	< 0.0001
Right Ventricle	0.019 [0.016–0.024]	0.016 [0.012–0.022]	< 0.0001
Right Atrium	0.014 [0.011–0.018]	0.011 [0.009–0.014]	< 0.0001
Short axis view:
Left Ventricle	0.013 [0.010–0.019]	0.013 [0.010–0.017]	0.41
Right Ventricle	0.035 [0.025–0.042]	0.036 [0.028–0.050]	0.003
Dice Metric
Long axis view:
Left Ventricle	0.978 [0.973–0.983]	0.986 [0.982–0.988]	< 0.0001
Right Ventricle	0.981 [0.976–0.984]	0.984 [0.978–0.988]	< 0.0001
Right Atrium	0.986 [0.983–0.989]	0.989 [0.985–0.991]	< 0.0001
Short axis view:
Left Ventricle	0.987 [0.982–0.991]	0.987 [0.983–0.990]	0.45
Right Ventricle	0.965 [0.958–0.975]	0.964 [0.951–0.972]	0.002

Comparison between the segmentation accuracy (percent variation and Dice metric) between U-Net based segmentation models trained entirely on synthetic frames generated by the generative adversarial network (PG GAN) and those trained on actual patient magnetic resonance imaging (MRI) frames. p-values were calculated using a paired non-parametric test

ISSN: 1471-2342