DUGCN: Deep Unfolding Gradient Consistency Network for fast MRI reconstruction

Compressed Sensing (CS) theory is a powerful technique to recover signal from highly under-sampled acquisition. It has been widely applied in fast Magnetic Resonance Imaging (MRI) reconstruction, effectively reducing MRI scan times and improving the patient experience. However, existing CS-MRI methods typically focus on minimizing the pixel-wise reconstruction error, and neglect the MR image characteristics, failing to restore high-frequency information. In this paper, we propose an efficient and reliable Deep Unfolding Gradient Consistency Network for fast MRI reconstruction, dubbed as DUGCN. First, to solve the ill-posed inverse problem, we consider a novel generalized CS energy model with implicit regularization by introducing a gradient consistency term, which encourages the model to ensure high-frequency consistency and pay more attention to delicate structure details. Then, we unfold the novel CS energy model using Half-Quadratic-Splitting algorithm and obtain four iterative sub-problems. We further develop corresponding deep architectures to effectively solve these subproblems with learnable parameters, and construct a end-to-end trainable DUGCN that combines the interpretable model-based methods and data-driven learning-based methods. Extensive experiments on various sampling patterns and acceleration rates demonstrate that our DUGCN can achieve high-quality MRI reconstruction with more accurate details, and outperforms the state-of-the-art methods.