WGAN for super-resolution megnatic resonance imaging

Magnetic resonance imaging (MRI), a tomography method, is widely used in clinical medicine for diagnosis of disease state．However, raw images often need to be further enhanced for better resolution．For super-resolution reconstruction, we propose a residual U-net WGAN back-sampling and super-sampling model, with perceptual, texture and adversarial loss function．A comparative experiment was conducted on 3000 MRI images, resulting in averaged PSNR of 33.09 and SSIM of 0.95, improving PSNR by 4.09 and SSIM by 0.06, proving the learning validity of low to high resolution images．This algorithm is stable, robust and could be widely used．