An Efficient Light-weight Network for Fast Reconstruction on MR Images

Background: In recent years, Deep Learning (DL) algorithms have emerged endlessly and achieved impressive performance, which makes it possible to accelerate Magnetic Resonance (MR) image reconstruction with DL instead of Compressed Sensing (CS) methods. However, a DL-based MR image reconstruction method has always suffered from its heavy learning parameters and poor generalization ability so far. Therefore, an efficient, light-weight network is still in desperate need of fast MR image reconstruction. Methods: We propose an efficient and light-weight MR reconstruction network (named RecNet) that uses a Convolutional Neural Network (CNN) to fast reconstruct high-quality MR images. Specifically, the network is composed of cascade modules, and each cascade module is further divided into feature extraction blocks and a data consistency layer. The feature extraction block can not only effectively extract the features of MR images, but also do not introduce too many parameters for the whole network. To stabilize the training procedure, the correction information of image frequency is adopted in the Data Consistency (DC) layer. Results: We have evaluated RecNet on a public dataset and the results show that the image quality reconstructed by RecNet is the best on the Peak Signal-To-Noise Ratio (PSNR) and structural similarity index (SSIM) evaluation standards. In addition, the pre-trained RecNet can also reconstruct high-quality MR images on an unseen dataset. Conclusion: The results demonstrate that the RecNet has superior reconstruction ability in various metrics than comparative methods. The RecNet can quickly generate high-quality MR images in fewer parameters. Furthermore, the RecNet has an excellent generalization ability on pathological images and different sampling rates data.