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

TL;DR

This work tackles the challenge of long MRI acquisition times by introducing DMSM, a Dual-domain Multi-path Self-supervised Diffusion Model. It combines dual-domain self-supervised training with a lightweight LHAN-DC reconstruction backbone and a multi-path inference strategy to produce high-quality reconstructions from undersampled data while providing uncertainty estimates. The method achieves state-of-the-art performance on fastMRI and IXI across multiple contrasts and acceleration factors, approaching fully supervised results with a compact ~0.8M parameter footprint and without requiring fully sampled training data. The uncertainty maps derived from multi-path reconstructions offer clinically interpretable guidance, potentially improving diagnostic confidence and workflow efficiency.

Abstract

Magnetic resonance imaging (MRI) is a vital diagnostic tool, but its inherently long acquisition times reduce clinical efficiency and patient comfort. Recent advancements in deep learning, particularly diffusion models, have improved accelerated MRI reconstruction. However, existing diffusion models' training often relies on fully sampled data, models incur high computational costs, and often lack uncertainty estimation, limiting their clinical applicability. To overcome these challenges, we propose a novel framework, called Dual-domain Multi-path Self-supervised Diffusion Model (DMSM), that integrates a self-supervised dual-domain diffusion model training scheme, a lightweight hybrid attention network for the reconstruction diffusion model, and a multi-path inference strategy, to enhance reconstruction accuracy, efficiency, and explainability. Unlike traditional diffusion-based models, DMSM eliminates the dependency on training from fully sampled data, making it more practical for real-world clinical settings. We evaluated DMSM on two human MRI datasets, demonstrating that it achieves favorable performance over several supervised and self-supervised baselines, particularly in preserving fine anatomical structures and suppressing artifacts under high acceleration factors. Additionally, our model generates uncertainty maps that correlate reasonably well with reconstruction errors, offering valuable clinically interpretable guidance and potentially enhancing diagnostic confidence.

Figures (8)

• Figure 1: Overview of the Dual-Domain Multi-Path Self-Supervised Diffusion Model (DMSM). During training, the original under-sampled k-space $y_u$ is randomly divided into 2 partitions $y_{u,p1}$ and $y_{u,p2}$. The training process takes $y_u,y_{u,p1},y_{u,p2}$ as inputs, and the self-supervised loss is performed on image domain $\mathcal{L}_{IC}$ and k-space domain $\mathcal{L}_{KC}$ for the diffusion model. Once trained, the diffusion model reconstructs MR images multiple times to get the final output and uncertainty.
• Figure 2: Architecture of Light-weighted Hybrid Attention Network (LHAN) in DMSM (Figure \ref{['fig:network']}). It consists of multiple Parameter-free Attention Blocks (PABs) and a Cross-Attention Transformer Block (CATB). The PABs enable MRI reconstruction feature extraction, while CATB fuse time index and the extract feature. The output of LHAN is then inputted into the DC layer.
• Figure 3: T1(top) and T2(bottom) MRI reconstruction results on fastMRI dataset across all performed baseline methods. Two different acceleration settings (R=4 and R=8) are included. The corresponding error maps are shown right below each reconstruction visualization. Closer to white indicates a better reconstruction compared to the ground truth. PSNR and SSIM values are also reported on the top of the reconstruction results.
• Figure 4: PD MRI reconstruction results on IXI dataset across all performed baseline methods. 4$\times$ acceleration results are presented. The corresponding error maps are shown right below each reconstruction. Closer to white indicates a better reconstruction compared to the ground truth. PSNR and SSIM values are also reported on the top of the reconstruction results.
• Figure 5: Uncertainty estimation on the multi-path averaged reconstruction. A result of FLAIR contrast on the fastMRI dataset with 4$\times$ acceleration is shown. Brighter values indicate higher values of variance and bias.