Mamba2SVN: a Mamba2 and reconstruction-cooperative sensitivity refinement-based variational network for parallel MRI reconstruction.
Authors
Affiliations (6)
Affiliations (6)
- Kunming University of Science and Technology, the Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming 650500, China, Kunming, Yunnan, 650500, China.
- Kunming University of Science and Technology, the Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming 650500, China, Kunming, 650500, China.
- The Third Affiliated Hospital of Kunming Medical University, Yunnan Cancer Hospital, the Third Affiliated Hospital of Kunming Medical University, Kunming, Yunnan, 650118, China.
- Shenzhen University, Shenzhen University 518037, Shenzhen, Guangdong, 518037, China.
- Xidian University, Xidian University 710000, Xi'an, Shaanxi, 710000, China.
- Tianjin University, Tianjin University, Tianjin 300072, Tianjin, 300072, China.
Abstract
Magnetic resonance imaging (MRI) is widely used for its excellent soft-tissue contrast and non-ionizing nature, but its long acquisition time remains a major bottleneck. In this work, we propose Mamba2SVN, a variational network that couples a Mamba2-based reconstruction branch with reconstruction-cooperative sensitivity refinement. First, we design a Mamba2-based Cross-iteration Feature Reconstruction (MCFR) module, where MCFRNet integrates Mamba2 blocks into a U-shaped backbone and incorporates a cross-iteration feature fusion (CIFF) mechanism. A preprocessing gated fusion (PPGF) unit adaptively fuses current bottleneck features with prior CIFF features, and wavelet-based preprocessing and postprocessing units (PRU/POU) balance reconstruction accuracy and computational efficiency. Second, we introduce a reconstruction-cooperative sensitivity refinement (RCSR) module that iteratively refines latent sensitivity variables by leveraging both previously reconstructed multi-coil k-space data and prior estimates of these variables, thereby reducing dependence on autocalibration signal (ACS) size. Third, we impose a dual data-consistency (DDC) mechanism that employs two complementary data-consistency operators at each iteration to alleviate error accumulation. Extensive experiments on multi-coil knee, brain, and cardiac datasets with various undersampling patterns demonstrate that Mamba2SVN consistently outperforms state-of-the-art deep learning-based methods. Mamba2SVN also remains robust when the ACS region is very small, suggesting its potential for highly accelerated clinical MRI protocols.