A Very Deep Densely Connected Network for Compressed Sensing MRI

Abstract

Convolutional neural network (CNN) has achieved great success in the compressed sensing-based magnetic resonance imaging (CS-MRI). Latest deep networks for CS-MRI usually consist of a stack of sub-networks, each of which refines the former image prediction to a more accurate one. However, as the sub-network number increases, the information in prior sub-networks has a little influence on subsequent ones, which increases the training difficulties and limits the reconstruction performance of the deep model. In this paper, we propose a novel network, named very deep densely connected network (VDDCN), for CS-MRI. Dense connections are introduced to connect any two sub-networks of VDDCN, so each sub-network can make full use of all former predictions, boosting the reconstruction performance of the whole network. The sub-network of VDDCN is composed of feature extraction and fusion block (FEFB) processing data in the image domain and data consistency (DC) layer enforcing the data fidelity in k-space. Specifically, in FEFB, multi-level features are extracted by the recursive feature extraction and fusion sub-blocks (RFEFSBs) and fused locally to obtain the compact features. The VDDCN is much deeper than the prior deep learning models and able to discover more MR image details. The experimental results demonstrate that our proposed VDDCN outperforms other state-of-the-art CS-MRI methods.