Reconstruction-Guided Multi-Stage Network For MRI Super-Resolution

Deep learning has provided an excellent solution for the MR image super-resolution. However, the existing methods mainly have two problems. Firstly, a single super-resolution network makes it difficult to accurately reconstruct lost high-frequency details. Secondly, a single-stage super-resolution network makes it difficult to achieve higher multiples of upsampling. Therefore, we propose a reconstruction-guided multi-stage network (RGMSNet) for MRI super-resolution. On the one hand, the guidance from the reconstruction prior can provide more clear features for super-resolution tasks. On the other hand, multi-stage supervision is introduced for gradual upsampling to reduce the difficulty of super-resolution. The reconstruction and super-resolution modules in our RGMSNet promote each other. Experimental results demonstrate the superiority of RGMSNet.