HiFi-Mamba: Dual-Stream W-Laplacian Enhanced Mamba for High-Fidelity MRI Reconstruction

TL;DR

HiFi-Mamba addresses the critical MRI reconstruction challenge of recovering high-fidelity images from undersampled k-space by introducing a frequency-aware dual-stream architecture that decouples low- and high-frequency information with a novel WL-based spectral decomposition. The HiFi-Mamba unit leverages cross-frequency guidance and spatially-aware parameter refinement to model global anatomical structure while preserving fine details, all within a streamlined unidirectional scanning pipeline to reduce redundancy. Empirical results on fastMRI and CC359 demonstrate state-of-the-art reconstruction accuracy with favorable efficiency, outperforming CNN-, Transformer-, and prior Mamba-based models across multiple metrics and acceleration factors. This work advances frequency-structured, efficiency-aware modeling for MRI reconstruction with practical implications for faster, higher-fidelity imaging.

Abstract

Reconstructing high-fidelity MR images from undersampled k-space data remains a challenging problem in MRI. While Mamba variants for vision tasks offer promising long-range modeling capabilities with linear-time complexity, their direct application to MRI reconstruction inherits two key limitations: (1) insensitivity to high-frequency anatomical details; and (2) reliance on redundant multi-directional scanning. To address these limitations, we introduce High-Fidelity Mamba (HiFi-Mamba), a novel dual-stream Mamba-based architecture comprising stacked W-Laplacian (WL) and HiFi-Mamba blocks. Specifically, the WL block performs fidelity-preserving spectral decoupling, producing complementary low- and high-frequency streams. This separation enables the HiFi-Mamba block to focus on low-frequency structures, enhancing global feature modeling. Concurrently, the HiFi-Mamba block selectively integrates high-frequency features through adaptive state-space modulation, preserving comprehensive spectral details. To eliminate the scanning redundancy, the HiFi-Mamba block adopts a streamlined unidirectional traversal strategy that preserves long-range modeling capability with improved computational efficiency. Extensive experiments on standard MRI reconstruction benchmarks demonstrate that HiFi-Mamba consistently outperforms state-of-the-art CNN-based, Transformer-based, and other Mamba-based models in reconstruction accuracy while maintaining a compact and efficient model design.

Figures (3)

• Figure 1: Illustration of scanning and decoupling. (a) Scanning strategies. Multi-directional scanning introduces redundancy, while the unidirectional approach avoids repeated access. Colors denote scan orders; red dashed boxes highlight redundant regions. (b) Visualization of k-space before and after the $\mathcal{w}$-Laplacian decomposition. Subfigures (b3) and (b4) show only the output branch retained for Mamba. The red circle marks the theoretical boundary between low- and high-frequency regions in k-space. This retained branch exhibits a cleaner, concentrated low-frequency spectrum and is better aligned with Mamba’s global modeling needs.
• Figure 2: Overview of the proposed HiFi-Mamba architecture. (a) The HiFi-Mamba Unit splits the input into high- and low-frequency components via the $\mathcal{w}$-Laplacian Block, processes them using the HiFi-Mamba Block and CRM (Condition Refinement Module), and fuses them with DSFA (Dual-Stream Fusion Attention). (b) The data consistency block. (c) CRM performs cross-resolution feature transformation. (d) The HiFi-Mamba block models frequency-aware sequences using Mamba-based token mixing. (e) The $\mathcal{w}$-Laplacian block performs idelity-preserving spectral decoupling. (f) DSFA fuses dual-frequency streams with adaptive weighting.
• Figure 3: Qualitative comparison on the fastMRI and CC359 datasets under single-coil settings. (a) Reconstruction results on the fastMRI knee dataset with acceleration factors $\text{AF}=4$ and $\text{AF}=8$. (b) Reconstruction results on the CC359 brain dataset under the same acceleration factors. The second row of each subplot shows the corresponding error maps. The blue boxes, yellow ellipses and red arrow highlight the details in the reconstruction results.