Fig. 6 | Scientific Reports

Fig. 6

From: Performance of a GPU- and time-efficient pseudo-3D network for magnetic resonance image super-resolution and motion artifact reduction

Fig. 6

Performance comparison of UNet and TS-RCAN for restoration of image quality under the most severe simulated motion artifact conditions (\({T}_{s}=9EG/18EG\)). Here, \(EG\) (echo group) denotes a fixed number of k-space lines, and \({T}_{s}\) represents the inter-motion interval during which the object remains stationary. Smaller \({T}_{s}\) values (fewer \(EGs\)) indicate shorter stationary periods, longer relative motion duration, and a larger proportion of k-space lines affected by motion, resulting in more severe artifacts. TS-RCAN with \(M=3\) significantly improved image quality in all cases, which was not the case for UNet and TS-RCAN with \(M=1\). Moreover, TS-RCAN outperformed UNet in SSIM/PSNR with improvement in all directions and severities of motion artifacts. (a)/(c) PSNR and SSIM of MA-corrupted and reduced images with 5 degrees in-plane rotation; (b)/(d) PSNR and SSIM of MA-corrupted and reduced images with 5 degrees in-plane rotation and 5 degrees through-plane rotation. ‘ + ’ represents the MAR with self-ensemble. * p < 0.05 and ** p < 0.0005.

Back to article page