FusionNet: A Parallel Interactive Neural Network for Compressed Sensing MRI Reconstruction

Compressed sensing provides the theoretical foundation for magnetic resonance imaging (MRI) reconstruction with undersampled k-space data with a sampling rate much less than the one required by the Nyquist-Shannon sampling theorem. However, CS-MRI principally relies on iterative numerical solvers which usually suffer from expensive computation cost and accurate handcrafted priori. In this paper, inspired by the popularity of deep learning, we propose a novel cascaded convolution neural network structure, called Fusion-Net, to accelerate MRI reconstruction. Different from other existing methods, our proposed FusionNet contains two parallel and interactive branches simultaneously performing on k-space and spatial-domain data. The experimental results show that the proposed method not only achieves competitive performance several state-of-the-art methods, but also outperforms other deep learning methods in terms of model scale and computational cost.