Accelerating MR imaging using deep learning on spatial frequency data
Abstract
Since the invention of the MRI machine, slow signal acquisition has been a significant barrier to the MRI process in getting a high throughput. A parallel imaging technique is used to simultaneously reconstruct a final MR image after acquiring fewer raw data with multi-RF coils to overcome this problem. However, the parallel imaging still cannot accelerate enough to reduce the overall time significantly. Therefore, the deep convolution neural network is presented to reconstruct a high-quality MR image with higher acceleration factors in this research. The proposed Multi-Level Pooling Encoder-Decoder Network (MLPED Net) performs the reconstructions by encoding the acquired signals. Then the crucial part of the network, a Multi-Level Pooling module (MLP), is applied to filter irrelevant signals, followed by a decoder module to reconstruct the MR images. The proposed network can be trained end-to-end with a newly presented loss function, Dual-Domain Loss (DDL). The experimental results are based on the published fastMRI knee dataset and show that the proposed MLPED Net outperforms the competing methods for both 4- and 8-fold accelerations.