Guided MRI Reconstruction via Schrödinger Bridge

TL;DR

The paper tackles the challenge of fast MRI reconstruction under undersampling by leveraging cross-contrast priors through a pixel-level translation framework based on the Schrödinger Bridge. It introduces I^2SB-Inversion, an SB-based guided reconstruction with an Inversion step that aligns guidance and target images to mitigate inter-modality misalignment, and it employs a CG-based data-consistency correction within the SB trajectory. The method demonstrates superior reconstruction quality and robustness on paired T1/T2 knee and brain datasets, achieving up to 14.4x acceleration compared with state-of-the-art guided and direct reconstruction methods. This work highlights the practicality of pixel-level cross-modality guidance and inversion in MRI reconstruction, offering improved structural fidelity and resilience to misalignment at high accelerations.

Abstract

Magnetic Resonance Imaging (MRI) is an inherently multi-contrast modality, where cross-contrast priors can be exploited to improve image reconstruction from undersampled data. Recently, diffusion models have shown remarkable performance in MRI reconstruction. However, they still struggle to effectively utilize such priors, mainly because existing methods rely on feature-level fusion in image or latent spaces, which lacks explicit structural correspondence and thus leads to suboptimal performance. To address this issue, we propose $\mathbf{I}^2$SB-Inversion, a multi-contrast guided reconstruction framework based on the Schrödinger Bridge (SB). The proposed method performs pixel-wise translation between paired contrasts, providing explicit structural constraints between the guidance and target images. Furthermore, an Inversion strategy is introduced to correct inter-modality misalignment, which often occurs in guided reconstruction, thereby mitigating artifacts and improving reconstruction accuracy. Experiments on paired T1- and T2-weighted datasets demonstrate that $\mathbf{I}^2$SB-Inversion achieves a high acceleration factor of up to 14.4 and consistently outperforms existing methods in both quantitative and qualitative evaluations.

Figures (7)

• Figure 1: (a) Forward step: The target image $\mathbf{x}_0$ gradually transforms into the guidance image $\mathbf{x}_N$, forming the SB trajectory $\{\mathbf{x}_0, \dots, \mathbf{x}_N\}$. Paired data $(\mathbf{x}_n$, $\mathbf{x}_0)$ are then used to train the denoiser $\epsilon_\theta$. (b) Backward step: Starting from $\mathbf{x}_N$, the target image $\mathbf{x}_0$ is reconstructed through iterative sampling, with data consistency enforced at each step to match the acquired k-space data. (c) Inversion step: After (b), $\mathbf{x}_0$ is converted into the aligned guidance image $\hat{\mathbf{b}}$, followed by resampling to obtain the final reconstruction $\hat{\mathbf{x}}_0$.
• Figure 2: Guided Reconstruction results in knee dataset at R = 11.2 and R = 14.4 . The top row shows the ground truth and the reconstructions obtained using different methods. The second row shows an enlarged view of the ROI, and the third row displays the error map of the reconstructions. Regions with blurring or artifacts in the compared methods are marked with yellow arrows.
• Figure 3: Comparison between direct reconstruction methods and the proposed $\mathrm{I}^2$SB-Inversion (with guidance) on the brain dataset at R = 11.2. The top row shows the ground truth and the reconstructions obtained using different methods. The second row shows an enlarged view of the ROI, and the third row displays the error map of the reconstructions. Regions with blurring or artifacts in the compared methods are marked with yellow arrows.
• Figure 4: Comparison between direct reconstruction methods and the proposed $\mathrm{I}^2$SB-Inversion (with guidance) on the brain dataset at R = 14.4. The top row shows the ground truth and the reconstructions obtained using different methods. The second row shows an enlarged view of the ROI, and the third row displays the error map of the reconstructions. Regions with blurring or artifacts in the compared methods are marked with yellow arrows.
• Figure 5: Reconstruction results under in-plane misalignment at an acceleration factor of R = 14.4. The top row shows the ground-truth and reconstructed images using $\mathrm{I}^2$SB-Recon and $\mathrm{I}^2$SB-Inversion methods w/wo misalignment. The bottom row displays enlarged ROIs for detailed comparison, with quantitative metrics reported in yellow and artifacts highlighted by yellow arrows.