Dual-Domain Self-Consistency-Enhanced Deep Unfolding Network for accelerated MRI reconstruction

Combining iterative optimization algorithms with deep neural networks, the deep unfolding method has presented great potential for accelerated Magnetic Resonance Imaging (MRI) reconstruction. The mainstream paradigm for deep unfolding methods is to apply the data fidelity term to the reconstructed k-space and the deep regularizer to the reconstructed image. However, this paradigm has the following limitations: (1) The traditional data fidelity term corrects only the data at sampled locations in k-space, leaving the data at unsampled locations outdated, which disrupts k-space self-consistency. (2) The regularizer relies solely on the limited prior knowledge from the image domain, lacking additional information and resulting in suboptimal detail reconstruction.

To address these issues, we propose a Dual-Domain Self-Consistency-Enhanced Deep Unfolding Network (DSDUN) for accelerated MRI reconstruction. Specifically, we additionally introduce a data fidelity term on the reconstructed image to indirectly update the entire k-space matrix by minimizing the total pixel error, thereby improving k-space self-consistency. Besides, we employ deep regularizers in both the image domain and frequency domain to exploit comprehensive information for better detail recovery. Furthermore, the dual-domain regularizers are tailored to the inherent characteristics of each respective domain, facilitating the effective extraction of domain-specific information.

Extensive experiments are conducted on three datasets under different undersampling patterns and acceleration factors. The experimental results show that the proposed DSDUN significantly improves MRI reconstruction quality (3.32 improvements in log frequency distance over the baseline and 14.64 improvements in peak signal to noise ratio).

In this paper, we propose an extended deep unfolding method to enhance k-space self-consistency and detail reconstruction. The experimental results highlight the significant potential of our method in accelerated MRI reconstruction.