Comparison of Reconstruction Performances Between Deep Learning and Iterative Methods in MR Compressed Sensing

Abstract

AbstractA deep residual learning convolutional neural network (CNN) is applied to improve image quality and speed up the reconstruction of compressed sensing magnetic resonance imaging. The reconstruction performances of proposed method are compared with iterative reconstruction methods.The proposed method adopts a deep residual learning CNN to learn the residual component between the input and output images (i.e., aliasing artifacts) for image reconstruction. The CNN-based reconstruction is compared with iterative reconstruction methods to clarify its characteristics. The peak-signal-to-noise ratio (PSNR) and the structural similarity index (SSIM) are examined for various numbers of sampling rates.The experimental results demonstrate that reconstruction time is drastically reduced to 0.022 s per image compared with that for conventional iterative reconstruction. Compared with iterative reconstruction, superior quality was obtained for low sampling rates (30% or 40%). These results suggest that CNN can learn the rules of the occurrence pattern of aliasing artifact. With the proposed method, the structure and small contrast in the original image are well preserved in the reconstructed image and aliasing artifacts are mostly removed.A deep residual learning CNN can recognize and predict aliasing artifacts. It is demonstrated that reconstruction time is significantly reduced.KeywordsCompressed sensingReconstructionDeep learning