⚡ electrical engineering

Dual-domain Multi-path Self-supervised Diffusion Model for Accelerated MRI Reconstruction

The paper proposes the Dual-domain Multi-path Self-supervised Diffusion Model (DMSM), a novel framework that improves accelerated MRI reconstruction by utilizing a self-supervised training scheme, a lightweight hybrid attention network, and a multi-path inference strategy to enhance accuracy, efficiency, and clinical interpretability through uncertainty estimation.

Original authors: Yuxuan Zhang, Jinkui Hao, Bo Zhou

Published 2026-04-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yuxuan Zhang, Jinkui Hao, Bo Zhou

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

Imagine you are trying to listen to a beautiful symphony being played in a room where someone is constantly making loud, distracting noises. You can only hear fragments of the music through the chaos.

In the world of medicine, MRI scans are like that symphony—they provide incredibly detailed "music" (images) of your body. However, to get a perfect, crystal-clear recording, the patient has to sit perfectly still for a long time. This is hard for kids, elderly patients, or anyone in pain. To speed things up, doctors use "accelerated" scans, which are like listening to the symphony through a heavy curtain: you get the gist, but the details are blurry and full of "noise" (artifacts).

This paper introduces a new AI system called DMSM that acts like a world-class conductor who can reconstruct that missing music perfectly, even when the recording is terrible.

Here is how it works, broken down into three "superpowers":

1. The "Self-Teaching" Student (Dual-Domain Self-Supervision)

The Problem: Most AI models are like students who need an "Answer Key" (fully completed, high-quality scans) to learn. But in a real hospital, you often don't have the Answer Key because the goal is to avoid the long, slow scan in the first place!

The DMSM Solution: Imagine a student who is given a jigsaw puzzle, but some pieces are missing. Instead of needing the picture on the box to learn, the student learns by looking at the pieces they do have and realizing, "If I put these two pieces together, they should match both in shape AND in color."

DMSM looks at the data in two ways at once: the "Image Domain" (the picture itself) and the "K-Space Domain" (the raw mathematical frequencies). By making sure the "picture" and the "math" agree with each other, the AI teaches itself how to fill in the blanks without ever needing a perfect "Answer Key."

2. The "Lightweight Athlete" (Hybrid Attention Network)

The Problem: Current high-end AI models are like massive, heavy sumo wrestlers. They are powerful, but they are slow, require huge amounts of energy (computing power), and take forever to move.

The DMSM Solution: DMSM is more like a lightweight Olympic sprinter. The researchers designed a "Hybrid Attention Network." Instead of using massive, heavy mathematical tools to look at every single pixel, it uses a clever, "parameter-free" trick to focus only on the important parts—like the sharp edges of an organ or the texture of a tissue. It gets the same high-quality results but does it much faster and with much less "brain power" (memory).

3. The "Second Opinion" (Multi-Path Uncertainty Estimation)

The Problem: In medicine, "I don't know" is a very important answer. Most AI models are "overconfident"—they give you an image and act like it’s 100% correct, even if they are actually just guessing.

The DMSM Solution: DMSM uses a "Multi-Path" strategy. Instead of drawing the image once, it draws it 15 different times, slightly varying its approach each time.

If all 15 drawings look exactly the same, the AI is saying: "I am very confident; this is definitely what the anatomy looks like."
If the 15 drawings look wildly different, the AI produces an "Uncertainty Map" (a heat map). It’s like the AI saying: "I think this is the brain, but I'm a bit shaky on this specific edge—Doctor, you might want to double-check this spot."

The Bottom Line

This paper provides a way to get fast, high-quality MRI scans that don't require perfect data to learn, don't require a supercomputer to run, and—most importantly—tell the doctor when they aren't quite sure about a detail. It’s a smarter, faster, and more honest way to look inside the human body.

Technical Summary: Dual-domain Multi-path Self-supervised Diffusion Model (DMSM)

1. Problem Statement

Accelerated Magnetic Resonance Imaging (MRI) is essential for improving clinical throughput and patient comfort by reducing scan times. However, traditional acceleration via k-space undersampling leads to aliasing artifacts. While Deep Learning (DL) and Diffusion Models have shown promise in addressing this, existing approaches face three critical bottlenecks:

Data Dependency: Most high-performing diffusion models require supervised learning, meaning they need fully-sampled (ground truth) datasets, which are difficult and time-consuming to acquire in clinical settings.
Computational Complexity: Current diffusion models often use heavy Transformer architectures with quadratic self-attention complexity, making them computationally expensive and slow for real-time clinical use.
Lack of Explainability: Existing methods typically provide a single reconstruction without quantifying the reliability of the output, lacking the uncertainty estimation necessary for clinical decision-making.

2. Methodology

The authors propose DMSM, a framework designed to reconstruct high-quality MRI images using only undersampled data. The methodology is built on three pillars:

Dual-domain Self-supervised Training: To eliminate the need for fully-sampled data, the model employs a partition mask strategy. The input undersampled k-space ( $y_u$ $y_{u}$ ) is randomly divided into two disjoint partitions ( $y_{u,p1}$ $y_{u, p 1}$ and $y_{u,p2}$ $y_{u, p 2}$ ). The model is trained using a combined loss function consisting of:
- Image-domain loss ( $L_{IC}$ ): Ensures consistency between reconstructions derived from different partitions.
- K-space domain loss ( $L_{KC}$ ): Ensures the reconstructed k-space aligns with the original undersampled measurements.
- Diffusion loss ( $L_{DM}$ ): A standard denoising score-matching loss.
Lightweight Hybrid Attention Network (LHAN): To reduce the computational burden of Transformers, the authors designed a custom backbone. It utilizes Parameter-free Attention Blocks (PABs), which use symmetric activation functions ( $\sigma_a$ ) to generate attention maps without adding trainable parameters. This is followed by a Cross-Attention Transformer Block (CATB) that efficiently fuses time-step conditioning with spatial features. A Data Consistency (DC) layer is integrated to anchor the generative process to the physical k-space measurements.
Multi-path Inference Strategy: During inference, the model performs multiple sampling paths with different random noise initiations. The final reconstruction is the mean of these paths, while the standard deviation across paths serves as a pixel-wise uncertainty map.

3. Key Contributions

Self-supervised Paradigm: A novel training scheme that enables high-fidelity reconstruction using only undersampled data, making it highly practical for real-world clinical environments.
Efficiency-Performance Balance: The LHAN architecture achieves state-of-the-art results with a significantly smaller footprint (only 0.8M parameters) compared to heavy Transformer-based models.
Clinical Interpretability: The introduction of uncertainty estimation via multi-path sampling provides a mechanism for radiologists to identify potentially unreliable areas in a reconstructed image.

4. Results

The model was evaluated on the fastMRI and IXI datasets across various contrasts (T1, T2, FLAIR, PD) and acceleration factors ( $R=4, R=8$ ).

Quantitative Superiority: DMSM outperformed several supervised and self-supervised baselines (e.g., SSDiff, SSAD-MRI, and D5C5). For example, on T1-weighted fastMRI at $4\times$ acceleration, it achieved a PSNR of 39.15 dB, significantly higher than the previous best self-supervised method.
Robustness: The model maintained high performance even under high acceleration ( $R=8$ ), where other methods suffered from structural distortions or intensity drifts.
Uncertainty Correlation: The generated uncertainty maps showed a strong spatial correlation with actual reconstruction errors (measured via Pearson Correlation Coefficient), validating the reliability of the uncertainty estimation.
Efficiency: The model achieved these results with a compact parameter count, demonstrating a superior trade-off between accuracy and computational cost.

5. Significance

DMSM represents a significant step toward autonomous and reliable AI in medical imaging. By removing the requirement for paired ground-truth data, it lowers the barrier for deploying advanced generative models in hospitals. Furthermore, by providing uncertainty maps, the model moves beyond "black-box" reconstruction toward a more explainable AI (XAI) framework, which is crucial for building trust between AI systems and medical professionals in diagnostic workflows.