Deep residual network for highly accelerated fMRI reconstruction using variable density spiral trajectory

Compressed sensing has proved itself as a useful technique for accelerating time-consuming fMRI acquisition. However, its intrinsic iterative algorithm of solving optimization problems limits its practical usage. In addition, it may still suffer from residual errors and artifacts when more aggressive acceleration factors are adopted. Deep neural networks have recently shown great potential in computer vision and image processing. Nevertheless, few attempts have been made concerning fMRI reconstruction. In this paper, we propose a deep residual network for faster and better image reconstruction with 20x acceleration. The network is made up of various residual blocks, and is trained to identify the mapping relationship between an aliased image and a fully-recovered image. Mean square error criterion refined with data-consistency loss is employed to evaluate the ‘distance’ between the network output and ground truth. Results showed that the proposed method can achieve superior image quality and better preserves dynamic features than other state-of-art methods. In addition, the reconstruction can be extremely fast with a one-way deployment on a feed-forward network.