Recurrent Mamba for efficient and high-fidelity MRI reconstruction

Fast Magnetic Resonance Imaging (MRI) reconstruction is a fundamental task in medical imaging, where balancing computational efficiency and high-fidelity image restoration remains a critical challenge. Recent advances in State Space Models (SSMs), particularly Mamba, have demonstrated strong capabilities in modeling long-range dependencies while maintaining efficiency. However, directly applying Mamba to MRI reconstruction poses limitations, including local pixel forgetting and an inability to capture fine-grained structural details. To address these challenges, we propose Recurrent Mamba (RCM), a novel framework that integrates state-space modeling with recurrent structures for efficient and high-fidelity MRI reconstruction. At its core, RCM introduces the Recurrent Mamba Block (RCMB), a cyclic state-space module that employs a spatially aware 2D selective state-space mechanism (2D_SSM) to mitigate pixel forgetting and enhance local feature representation. Additionally, an Adaptive Convolution Block (ACB) is incorporated to optimize feature extraction while maintaining computational efficiency. Expanding on RCMB, we develop the Unfolded Recurrent Mamba Module (URCM), which hierarchically refines multi-scale features through iterative learning with Recurrent Learning Units (RLUs) and a Refinement Module (RM). Experimental results show that RCM achieves considerable performance improvements on four MRI datasets with different sequences, while maintaining excellent parameter efficiency, with only 1.03 M trainable parameters.