Fp^roi-GAN with Fused Regional Features for the Synthesis of High-Quality Paired Medical Images

Table 3

Results of the quantitative evaluation of the BRATS 2017 dataset (mean ± standard deviation), where we compare the measurements of the different synthesis methods over the whole image domain and the tumor domain at a significance level of 0.05, and the underline indicates that Fp^roi-GAN is statistically significantly different from the other methods.

Data	Region	Methods	PSNR	SSIM	MS-SSIM
HGG	Whole image	DCGAN [11]	25.749 ± 3.49	0.882 ± 0.04	0.890 ± 0.05
		Pix2Pix [15]	28.938 ± 4.68	0.952 ± 0.03	0.956 ± 0.05
		cycleGAN [16]	34.280 ± 4.85	0.984 ± 0.02	0.984 ± 0.05
		Fp^roi-GAN	34.884 ± 5.18	0.986 ± 0.02	0.987 ± 0.04
	Tumor region	DCGAN [11]	29.539 ± 5.05	0.903 ± 0.02	0.910 ± 0.05
		Pix2Pix [15]	33.031 ± 5.99	0.951 ± 0.02	0.952 ± 0.04
		cycleGAN [16]	35.652 ± 5.97	0.977 ± 0.03	0.970 ± 0.04
		Fp^roi-GAN	41.888 ± 6.06	0.997 ± 0.004	0.993 ± 0.03
LGG	Whole image	DCGAN [11]	23.093 ± 4.71	0.895 ± 0.11	0.908 ± 0.06
		Pix2Pix [15]	25.912 ± 4.95	0.933 ± 0.09	0.945 ± 0.07
		cycleGAN [16]	28.045 ± 4.47	0.958 ± 0.08	0.966 ± 0.03
		Fp^roi-GAN	30.044 ± 4.21	0.964 ± 0.08	0.974 ± 0.03
	Tumor region	DCGAN [11]	25.809 ± 4.39	0.892 ± 0.09	0.911 ± 0.07
		Pix2Pix [15]	30.228 ± 5.28	0.939 ± 0.08	0.948 ± 0.07
		cycleGAN [16]	29.192 ± 7.22	0.993 ± 0.01	0.983 ± 0.06
		Fp^roi-GAN	40.440 ± 7.51	0.997 ± 0.02	0.990 ± 0.03