HiFi-MambaV2: Hierarchical Shared-Routed MoE for High-Fidelity MRI Reconstruction

TL;DR

HiFi-MambaV2 addresses the challenge of reconstructing high-fidelity MRI from undersampled k-space by unifying frequency-aware decomposition with content-adaptive computation. It introduces SF-Lap for stable, alias-resistant multi-scale frequency separation and a hierarchical Shared-Routed MoE that performs per-pixel top-1 routing across frequency-specific experts, supplemented by a lightweight SE-guided global context path within an unrolled data-consistency framework. Across five public datasets and multiple acceleration factors, the method delivers state-of-the-art PSNR, SSIM, and NMSE, with robust performance in both single- and multi-coil settings and favorable efficiency compared to Transformer-based and prior Mamba baselines. Ablation studies validate the complementary contributions of SF-Lap, MR-MoE routing, and the global context, highlighting improved frequency modeling, anatomical coherence, and stable cross-depth behavior. These results suggest HiFi-MambaV2 offers a practical, scalable approach for high-fidelity MRI reconstruction with strong generalization and potential clinical impact.

Abstract

Reconstructing high-fidelity MR images from undersampled k-space data requires recovering high-frequency details while maintaining anatomical coherence. We present HiFi-MambaV2, a hierarchical shared-routed Mixture-of-Experts (MoE) Mamba architecture that couples frequency decomposition with content-adaptive computation. The model comprises two core components: (i) a separable frequency-consistent Laplacian pyramid (SF-Lap) that delivers alias-resistant, stable low- and high-frequency streams; and (ii) a hierarchical shared-routed MoE that performs per-pixel top-1 sparse dispatch to shared experts and local routers, enabling effective specialization with stable cross-depth behavior. A lightweight global context path is fused into an unrolled, data-consistency-regularized backbone to reinforce long-range reasoning and preserve anatomical coherence. Evaluated on fastMRI, CC359, ACDC, M4Raw, and Prostate158, HiFi-MambaV2 consistently outperforms CNN-, Transformer-, and prior Mamba-based baselines in PSNR, SSIM, and NMSE across single- and multi-coil settings and multiple acceleration factors, consistently surpassing consistent improvements in high-frequency detail and overall structural fidelity. These results demonstrate that HiFi-MambaV2 enables reliable and robust MRI reconstruction.

Figures (3)

• Figure 1: Each panel visualizes channel-wise 2D log-magnitude Fourier spectra (14×14 thumbnails) of intermediate feature maps; top row is before the block and bottom row is after. Conv Block preserves pronounced cross-shaped high-frequency responses, reflecting strong sensitivity to local edges and textures. Pure Mamba suppresses mid/high frequencies and concentrates energy near the DC component, evidencing a low-pass, global-context behavior. Our HiFi–Mamba block produces a balanced spectrum: low frequencies dominate (global consistency) while moderate high-frequency energy is retained (local detail). This frequency-domain complementarity explains the downstream gains we report—integrating Mamba with convolution unifies global semantics and local precision.
• Figure 2: Overview of the proposed HiFi-MambaV2 architecture. (a) The HiFi-Mamba Unit (×2 per Group) applies the Lightweight SE-Guided Global Context Path (LSGP) module for residual global enhancement, followed by frequency decomposition via the SF-Lap module to produce stable high- and low-frequency components. Each frequency stream is processed by frequency-aware HiFi-Mamba Blocks and routed through a hierarchical shared-routed MoE mechanism with four paths. (b) The Data Consistency (DC) block enforces fidelity with acquired k-space data after each Group. (c) The HiFi-Mamba Group (×8 cascades) stacks two Mamba Units and one DC block within an unrolled optimization pipeline. (d) Condition Refinement Module (CRM) performs cross-resolution feature transformation. (e) The LSGP enhances global anatomical awareness. (f) The HiFi-Mamba block models frequency-aware sequences using Mamba-based token mixing. (g) The Separable Frequency-Consistent Laplacian Pyramid (SF-Lap) achieves energy-preserving, alias-resistant multi-scale decomposition using depthwise binomial filtering and symmetric reconstruction. (h) The Dual-Frequency Spatial Attention (DFSA) fuses the routed experts’ outputs through adaptive weighting to form the final feature representation.
• 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 AF=4 and 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.