SGSR: Structure-Guided Multi-Contrast MRI Super-Resolution via Spatio-Frequency Co-Query Attention

TL;DR

SGSR tackles the problem of reconstructing high-resolution MRI from low-resolution, multi-contrast acquisitions by exploiting contrast-invariant structural information. It introduces a structure-guided multi-contrast SR framework based on co-query attention (CQA) that operates in both spatial and frequency domains: Spatial Co-Query Attention (SCQA) uses a shared structural query $Q$ and contrast-specific keys/values to refine structures via $\tilde{X}_c = \sigma\left(\frac{Q (K_c)^T}{\sqrt{d}}\right) V_c$, while Frequency Co-Query Attention (FCQA) applies localized 2D FFT-based tokens to separate high-frequency structure from low-frequency appearance and refine them in the Fourier domain. The backbone is an encoder-decoder network with four residual groups, trained with an $L_1$ loss, and experiments on fastMRI knee and M4Raw brain show SGSR surpasses state-of-the-art MCSR and SISR methods with significant PSNR/SSIM gains and lower compute/memory. The work demonstrates that explicitly modeling cross-contrast structural information via efficient CQA can yield robust, structure-consistent SR and suggests extension to other modalities such as MRI-CT.

Abstract

Magnetic Resonance Imaging (MRI) is a leading diagnostic modality for a wide range of exams, where multiple contrast images are often acquired for characterizing different tissues. However, acquiring high-resolution MRI typically extends scan time, which can introduce motion artifacts. Super-resolution of MRI therefore emerges as a promising approach to mitigate these challenges. Earlier studies have investigated the use of multiple contrasts for MRI super-resolution (MCSR), whereas majority of them did not fully exploit the rich contrast-invariant structural information. To fully utilize such crucial prior knowledge of multi-contrast MRI, in this work, we propose a novel structure-guided MCSR (SGSR) framework based on a new spatio-frequency co-query attention (CQA) mechanism. Specifically, CQA performs attention on features of multiple contrasts with a shared structural query, which is particularly designed to extract, fuse, and refine the common structures from different contrasts. We further propose a novel frequency-domain CQA module in addition to the spatial domain, to enable more fine-grained structural refinement. Extensive experiments on fastMRI knee data and low-field brain MRI show that SGSR outperforms state-of-the-art MCSR methods with statistical significance.

Figures (4)

• Figure 1: (a) Overall architecture of our SGSR; (b) Standard cross-attention; (c) The proposed co-query attention (CQA) within the spatial module.
• Figure 2: Frequency co-query attention (FCQA) module. The above process is followed by Inverse FFT to transform back to spatial domain.
• Figure 2: Ablation study on M4Rawlyu_m4raw_2023 4× task. * for values that the full version significantly outperforms with $p$-value $< 0.01$.
• Figure 3: Qualitative results on SR predictions and error maps of fastMRI zbontar_fastmri_2019 (top) and M4Rawlyu_m4raw_2023 (bottom). Metrics are shown in PSNR/SSIM for each image. Input (bottom) and ground truth (top) are shown in the first column.