⚡ electrical engineering

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

HiFi-Mamba is a novel dual-stream architecture that enhances high-fidelity MRI reconstruction by integrating W-Laplacian spectral decoupling to separately model low- and high-frequency features and employing a streamlined unidirectional traversal to eliminate scanning redundancy, thereby outperforming existing CNN, Transformer, and Mamba-based models in accuracy and efficiency.

Original authors: Hongli Chen, Pengcheng Fang, Yuxia Chen, Yingxuan Ren, Jing Hao, Fangfang Tang, Xiaohao Cai, Shanshan Shan, Feng Liu

Published 2026-05-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hongli Chen, Pengcheng Fang, Yuxia Chen, Yingxuan Ren, Jing Hao, Fangfang Tang, Xiaohao Cai, Shanshan Shan, Feng Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Blurry Snapshot"

Imagine you are trying to take a perfect photo of a busy city, but you only have time to snap a few quick, incomplete pictures. If you try to piece those fragments together, the result is usually a blurry, distorted mess with weird artifacts (like ghostly lines or missing buildings).

This is exactly what happens in MRI scans. To get a clear picture of the inside of your body, the machine needs to collect a massive amount of data (called "k-space"). However, collecting all that data takes a long time, which makes patients uncomfortable and causes them to move, ruining the scan. So, doctors often take "undersampled" data—less information to save time. The challenge is: How do we use AI to fill in the missing pieces and create a crystal-clear image from a blurry, incomplete one?

The Old Solutions: The "Local" and the "Global"

Before this paper, AI tried to solve this in two main ways, but both had flaws:

The "Local" Detective (CNNs): These models are like a detective who only looks at one brick at a time. They are great at seeing fine details (like the texture of a brick) but terrible at understanding the whole building's structure. They miss the big picture.
The "Global" Architect (Transformers): These models look at the whole building at once. They understand the structure perfectly but are incredibly slow and expensive to run, like trying to hire a team of 1,000 architects just to design a shed.
The "New Kid" (Mamba): Recently, a new type of AI called Mamba arrived. It's fast (like the detective) but can also see the whole building (like the architect). However, when researchers tried to use Mamba for MRI, it had three specific problems:
- It was too "lazy" with details: It ignored the high-frequency details (the sharp edges and tiny textures) needed for a clear medical image.
- It was redundant: It looked at the image from four different directions (up, down, left, right), doing the same work four times. It was like reading a book four times just to make sure you understood it.
- It was too local: It didn't pay enough attention to the neighbors of a specific pixel, which is crucial for smooth, natural-looking anatomy.

The Solution: HiFi-Mamba (The "Dual-Stream Chef")

The authors created HiFi-Mamba (High-Fidelity Mamba). Think of this as a two-kitchen restaurant designed to cook the perfect meal (the MRI image) efficiently.

1. The "Spectral Decoupling" (Separating the Soup from the Garnish)

The first step is a special tool called the w-Laplacian block. Imagine you have a bowl of soup. You want to separate the clear broth (the smooth, big structures of the body) from the crunchy garnish (the sharp edges and fine details).

Low-Frequency Stream (The Broth): This goes to the main Mamba chef. Since it's smooth and simple, the chef can easily understand the "big picture" and the overall shape of the anatomy.
High-Frequency Stream (The Garnish): This is kept separate. It contains all the sharp edges and tiny details that the main chef usually ignores.

2. The "Guided Chef" (Cross-Stream Guidance)

Here is the magic trick. The main chef (processing the "broth") is usually blind to the "garnish." In HiFi-Mamba, the chef gets a special note from the garnish station.

The "garnish" stream tells the "broth" stream: "Hey, there's a sharp edge right here!" or "Don't blur this part!"
This allows the fast Mamba model to build the big structure while knowing exactly where to keep the sharp details. It's like a painter who knows exactly where to put the fine lines because someone is whispering the coordinates to them.

3. The "One-Way Street" (Unidirectional Scanning)

Old Mamba models looked at the image from four directions (up, down, left, right), which was slow and repetitive.

HiFi-Mamba changes this to a single, streamlined path. It reads the image in one direction only.
Analogy: Imagine reading a book. The old way was reading the page left-to-right, then right-to-left, then top-to-bottom, then bottom-to-top. The new way is just reading it once, left-to-right, but reading it so well that you don't need to go back. This makes the process much faster and less redundant.

4. The "Smart Neighborhood Watch" (Spatially-Aware Refinement)

Standard Mamba sometimes treats every pixel as an island. HiFi-Mamba adds a step where the AI looks at its immediate neighbors (like a neighborhood watch) before making a decision. This ensures that the transition between different body parts (like skin to muscle) is smooth and natural, not jagged.

The Results: Faster, Sharper, Smarter

The paper tested this new "Dual-Stream Chef" against the best existing methods on standard MRI datasets (knee and brain scans).

Accuracy: HiFi-Mamba produced clearer images with fewer errors than the old "Local" detectives, the slow "Global" architects, and even other versions of Mamba. It was particularly good at preserving fine details (like the edges of bones) without creating fake structures.
Efficiency: Because it stopped looking at the image from four directions and started using a single, smart path, it used significantly less computing power.
- Analogy: It's like switching from a gas-guzzling truck that drives in circles to a sleek electric car that takes the most direct route. It gets you to the destination (a clear image) faster and uses less fuel (computer power).

Summary

HiFi-Mamba is a new AI architecture that fixes the flaws of previous models by:

Separating the "smooth" parts of the image from the "sharp" parts.
Letting the sharp parts guide the smooth parts so nothing gets blurred.
Reading the image in one efficient direction instead of four redundant ones.

The result is a system that creates high-quality, detailed MRI images much faster and with less computing power than ever before.

Technical Summary: HiFi-Mamba

Problem Statement

Reconstructing high-fidelity Magnetic Resonance Imaging (MRI) from undersampled k-space data is a critical challenge. While modern protocols accelerate scans by undersampling the frequency domain, this violates the Nyquist-Shannon sampling criterion, leading to aliasing artifacts and degraded image quality.

Existing deep learning approaches face specific limitations:

CNNs: While effective at modeling local features, their inherent locality limits the capture of long-range dependencies crucial for global anatomical structures.
Transformers: Although capable of global context modeling via self-attention, they often suffer from quadratic computational complexity, making them inefficient for high-resolution MRI. Variants like Swin Transformers improve efficiency but constrain the receptive field, re-introducing the trade-off between efficiency and global modeling.
Mamba (SSM): Recent State Space Models (SSMs) like Mamba offer linear-time scalability and strong sequence modeling. However, direct application to MRI reveals three key limitations:
1. Redundant Scanning: Existing vision Mamba variants (e.g., vMamba) use multi-directional scanning strategies that introduce significant computational redundancy.
2. Locality: Standard Mamba derives state-space parameters independently for each token via local transformations, neglecting neighboring context essential for coherent anatomical structures.
3. Frequency Insensitivity: Existing variants lack sensitivity to high-frequency components, which are vital for preserving fine anatomical details in MRI.

Methodology: HiFi-Mamba

The authors propose HiFi-Mamba (High-Fidelity Mamba), a dual-stream architecture designed to address the aforementioned limitations. The network follows an unrolled optimization framework with $K=8$ cascaded HiFi-Mamba Groups, each containing two Mamba Units and a Data Consistency (DC) block.

Core Architecture Components

1. The w-Laplacian (WL) Block: Spectral Decoupling
To enable frequency-aware processing, the input feature map is split into two streams using a lightweight w-Laplacian block. Unlike traditional Laplacian pyramids which are resolution-oriented, this block uses a wavelet-based formulation for structured, reversible frequency separation:

Low-Frequency Stream: Generated by upsampling the LL subband of a 2D Discrete Wavelet Transform (DWT). This stream provides a clean, concentrated low-frequency spectrum suitable for global modeling.
High-Frequency Stream: Defined as the residual between the refined input and the low-frequency base. This stream preserves fine anatomical details.

2. The HiFi-Mamba Block: Guided Global Modeling
This block processes the low-frequency stream ( $F_{low}$ ) but is explicitly conditioned by the high-frequency stream to overcome Mamba's insensitivity to fine details.

Cross-Frequency Guidance: An anatomical guidance map ( $G$ ) is extracted from the high-frequency stream via a Condition Refinement Module (CRM). This map modulates the state-space parameters ( $B$ and $C$ ) of the Mamba block. Specifically, gating mechanisms distill salient high-frequency information (e.g., edges, artifacts) from $G$ and inject them into the state transition and output equations. This ensures that global modeling is informed by local structural details.
Spatially-Aware Refinement: To address the lack of spatial awareness in standard Mamba, the state-space parameters ( $B, C, \Delta$ ) are refined using a 1D depth-wise convolution (kernel size $k=7$ ). This allows parameters for each token to be influenced by spatial neighbors, injecting local context into the global sequence modeling.
Unidirectional Traversal: To eliminate scanning redundancy, the block adopts a streamlined unidirectional traversal strategy, avoiding the repeated access inherent in multi-directional scanning.

3. Dual-Stream Fusion
The enhanced low-frequency features (from the HiFi-Mamba block) and the refined high-frequency features (processed via CRM) are concatenated and fused using a Dual-Stream Fusion Attention (DSFA) module. This module adaptively weights the two streams to produce the final fused feature map, which is then combined with the original input via a residual connection.

Key Contributions

Dual-Stream Architecture: A novel framework that decouples features into low- and high-frequency streams, allowing the Mamba block to focus on global low-frequency structures while selectively integrating high-frequency details via adaptive state-space modulation.
w-Laplacian Block: A fidelity-preserving spectral decoupling mechanism that replaces resolution-based hierarchies with wavelet-based separation, providing a dedicated low-frequency input aligned with Mamba's global modeling needs.
Efficient Scanning Strategy: The adoption of a unidirectional traversal strategy that maintains long-range modeling capabilities while significantly reducing computational redundancy compared to multi-directional scanning.
Cross-Frequency Guidance: A mechanism that modulates Mamba's state-space parameters using high-frequency anatomical cues, resolving the inherent locality and frequency insensitivity of standard Mamba designs.

Experimental Results

The model was evaluated on two public datasets: fastMRI (knee) and CC359 (brain), under acceleration factors (AF) of 4× and 8×.

Performance: HiFi-Mamba consistently outperforms state-of-the-art CNN-based (UNet, ISTA-Net), Transformer-based (ReconFormer, FPSFormer), and Mamba-based (LMO) baselines.
- On fastMRI (AF=8), HiFi-Mamba (P1) achieved 31.81 dB PSNR and 0.762 SSIM, surpassing the previous best Mamba model (LMO) which scored 31.10 dB PSNR.
- On CC359 (AF=8), it achieved 28.49 dB PSNR and 0.810 SSIM, again outperforming LMO (27.99 dB PSNR).
Efficiency: The unidirectional scanning and streamlined design yield significant computational savings.
- HiFi-Mamba (P2) achieved 31.38 dB PSNR with only 67.87G FLOPs, substantially lower than LMO (484.98G FLOPs) and Transformer-based models (e.g., ReconFormer at 270.60G FLOPs).
- Even the higher-performance P1 variant (270.37G FLOPs) remains more efficient than LMO while delivering superior accuracy.
Qualitative Analysis: Visual comparisons show HiFi-Mamba produces sharper contours and fewer hallucinated structures compared to Transformers, while preserving fine anatomical details better than CNNs and other Mamba variants.

Significance and Claims

The paper claims that HiFi-Mamba establishes a new state-of-the-art in MRI reconstruction by effectively balancing reconstruction fidelity and computational efficiency.

The authors argue that their work bridges the gap between current vision-based Mamba variants and the specific spectral characteristics of MRI. By introducing frequency-aware dual-stream processing and eliminating redundant scanning, HiFi-Mamba offers a promising direction for frequency-structured and efficiency-aware modeling. The results suggest that the proposed architecture is robust across different anatomical structures and acceleration factors, providing a holistic solution that avoids the trade-offs typically found between global modeling capability and computational cost.