MRI Reconstruction Using Graph Reasoning Generative Adversarial Network

The deep learning-based CS-MRI methods have been demonstrated to be able to reconstruct high-precision MR images. However, it can be observed that most current deep learning-based CS-MRI methods capture long-range dependencies by stacking multiple convolutional layers, which is computationally inefficient. The latent graph neural network has been proposed to efficiently capture long-range dependencies, which can address the above issue. Besides, there are very few works introducing graph neural networks (GNNs) into MRI reconstruction. In this paper, we propose a novel graph reasoning generative adversarial network, termed as GRGAN, by introducing the graph reasoning networks into MRI reconstruction, where the graph reasoning networks are embedded in the generator to capture long-range dependencies more efficiently. In addition, we propose the multi-scale aggregated residual blocks, termed as MARBs, and introduce them into the proposed GRGAN to extract multi-scale feature information effectively. The experimental results demonstrate that the proposed GRGAN surpasses the state-of-the-art deep learning-based CS-MRI methods with fewer model parameters.