MR-FusionMamba: A Visual Mamba Network with Range-Null Decomposition for Multi-Modal MRI reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accelerated multi-modal magnetic resonance imaging (MRI) reconstructs target images from undersampled data using a reference modality.
• Current deep learning methods like CNNs and Transformers have limitations in capturing global context or computational efficiency.

Purpose of the Study:

• To develop an efficient deep learning model for accelerated multi-modal MRI reconstruction.
• To address the limitations of CNNs and Transformers in feature extraction and global information modeling.

Main Methods:

• Proposed MR-FusionMamba, integrating Mamba blocks into dual U-shaped networks for dual-input support.
• Utilized the Range-Null Decomposition theorem to enhance data consistency.
• Evaluated performance on the BraTS dataset.

Main Results:

• MR-FusionMamba demonstrates efficient multi-modal MRI reconstruction.
• The model effectively captures global context with linear complexity, outperforming existing methods.
• Achieved state-of-the-art (SOTA) performance in experiments.

Conclusions:

• MR-FusionMamba offers a promising solution for accelerated multi-modal MRI reconstruction.
• The integration of Mamba architecture provides an efficient alternative to CNNs and Transformers.
• The method shows significant potential for advancing fast MRI techniques.