表 1.

不同方法在MR脑影像超分辨率任务上的测试结果

Metrics of different methods in super-resolution task for routine brain MRI

Item	Interpolation	FSRCNN	U-net	SRFBN	SCSR	MSCSR
FSRCNN: Fast super-resolution convolutional neural network, SRFBN: Super-resolution feedback Network, SCSR: Structure-constrained super-resolution network, MSCSR: Multi-modality structure-constrained super-resolution network. PSNR: Peak signal-to-noise ratio, SSIM: Structural similarity. GM: Gray matter, WM: White matter, CSF: Cerebrospinal fluid.
PSNR	14.47±0.60	19.39±0.75	21.83±0.78	24.40±0.68	26.34±0.89	33.11±0.86
SSIM	0.943±0.006	0.978±0.003	0.989±0.002	0.992±0.001	0.995±0.001	0.996±0.001
Volume Error
GM	19 622(3%)	68 610(10%)	209 725 (33%)	8372 (1%)	14 799 (2%)	9967(1%)
WM	84 461 (18%)	48 629(10%)	159 236 (34%)	9383 (2%)	6207 (1%)	8455 (2%)
CSF	98 681 (35%)	13 918(5%)	83 004 (30%)	7818(3%)	10 300 (3%)	24 187 (8%)