GL-mamba-net: A magnetic resonance imaging restoration network with global-local mamba

In clinical practice, magnetic resonance imaging is essential for disease diagnosis and evaluation, although it generally requires a prolonged scanning duration. Recently, several deep learning-based restoration techniques have been introduced to accelerate MRI acquisition. However, most existing MRI restoration methods struggle to fully capture local texture features and have limitations in fusing global and local features. To address this issue, the paper introduces a dual path Mamba network. This method uses a strategy of multi-scale local feature and global feature fusion for under-sampled single-coil image domain data, significantly improving both the quality and efficiency of image restoration. Specifically, 1) A multi-scale local Mamba block is proposed, which extracts local information from different regions through a multi-scale window mechanism, capturing diverse local features. 2) A new feature fusion block is proposed, which fuses global and local information to enhance the completeness of feature expression. 3) A dual path Mamba network architecture is proposed. This dual-path design significantly improves feature extraction capability and adaptability in complex and dynamic data environments. Through comprehensive experiments on the NAMIC, fastMRI and BraTS datasets, it is shown that the proposed network surpasses current leading methods across various evaluation metrics, particularly excelling in restoring texture details and tissue structures.