Abstract
MMR-Mamba is a novel framework for multi-modal MRI reconstruction that leverages the efficiency of State Space Models (Mamba) to fuse information across spatial and frequency domains. By combining a target-guided cross-modal mechanism with dual-domain fusion, it achieves state-of-the-art reconstruction quality while using significantly fewer parameters than Transformer-based alternatives.
Introduction
Magnetic Resonance Imaging (MRI) is a powerful diagnostic tool, but its clinical utility is often limited by slow acquisition speeds. To accelerate the process, clinicians often "under-sample" k-space data, which can lead to blurry images or significant artifacts. A common strategy to mitigate this is multi-modal reconstruction, which utilizes a fast-acquired, fully-sampled "reference" scan (such as a T1-weighted image) to assist in reconstructing a slow-acquired, under-sampled "target" scan (such as a T2-weighted image).
The technical challenge lies in effective feature fusion. While Convolutional Neural Networks (CNNs) are adept at capturing local textures, they often fail to model the long-range dependencies necessary for global structural integrity. Conversely, Transformers capture global context effectively but suffer from quadratic computational complexity ($O(N^2)$), making them computationally prohibitive for high-resolution medical imaging. MMR-Mamba addresses these limitations by utilizing the emerging Mamba architecture, which provides global receptive fields with linear complexity ($O(N)$).
Methods
MMR-Mamba introduces three primary novel components to optimize multi-modal fusion:
1. Target-modality guided Cross Mamba (TCM)
Located in the spatial domain, the TCM module utilizes Mamba blocks to extract features. It employs a unique gating mechanism where the target modality's features guide the integration of reference features. This ensures that the final reconstructed image retains the target's specific contrast characteristics rather than simply mimicking the reference scan.
2. Selective Frequency Fusion (SFF)
This module operates in the Fourier domain, treating the phase spectrum (which carries structural information) and the amplitude spectrum (which carries style and degradation information) differently. It features a Selective Amplitude Harmonization Module (SAHM) that dynamically adjusts and restores degraded amplitude features based on global statistics.
3. Adaptive Spatial-Frequency Fusion (ASFF)
ASFF serves as the bridge between domains. It utilizes Batch Normalization (BN) scaling factors as a proxy for channel importance. If a channel in the spatial domain is identified as having low informativeness, ASFF reinforces it with corresponding information from the frequency domain, and vice versa.
Datasets and Evaluation
The model was evaluated on two major benchmarks:
- BraTS Dataset: Brain MRIs using T1-weighted images as the reference to reconstruct T2-weighted target images.
- fastMRI Knee Dataset: Knee MRIs using Proton Density (PD) images as the reference to reconstruct Fat-Suppressed PD (FS-PD) target images.
Both datasets were tested at 4× and 8× acceleration factors to simulate clinical under-sampling scenarios.
Key Results
MMR-Mamba consistently outperformed seven state-of-the-art methods, including Pan-Mamba and SwinIR, across all metrics:
- Superior Accuracy: On the BraTS dataset (4× acceleration), MMR-Mamba achieved a PSNR of 40.98 dB, outperforming the second-best model by 2.14 dB.
- Extreme Efficiency: One of the most significant findings is the model's parameter efficiency. MMR-Mamba achieves these results with only 0.42 million parameters, whereas competing models like MTrans require 746 million parameters.
- Visual Fidelity: Error maps indicate that MMR-Mamba significantly reduces artifacts in complex anatomical regions, such as the lateral ventricles of the brain, where other models often exhibit ghosting or blurred edges.
Conclusion and Future Work
MMR-Mamba demonstrates that Mamba’s selective scanning mechanism is highly effective for medical imaging tasks. By selectively harmonizing amplitude in the frequency domain and using BN scaling for adaptive fusion, the framework preserves structural truth while effectively removing noise. Future research will focus on extending the model to 3D volumetric reconstruction and improving robustness to spatial misalignment between reference and target scans.
Summary updated