UniFS: Unified Multi-Contrast MRI Reconstruction via Frequency-Spatial Fusion

TL;DR

This work tackles the problem of generalizing multi-contrast MRI reconstruction across diverse k-space undersampling patterns without retraining. It introduces UniFS, a Unified Frequency-Spatial Fusion model that performs genuine frequency-domain fusion via Cross-Modal Frequency Fusion, Adaptive Mask-Based Prompt Learning, and Dual-Branch Complementary Refinement, coupled with a Joint Frequency-Spatial Reconstruction Loss. Experiments on BraTS and HCP show state-of-the-art performance across multiple acceleration factors and unseen patterns, evidencing strong cross-pattern generalization. The approach promises practical clinical impact by enabling robust, distribution-agnostic MCMR and paves the way for broader validation across centers and anatomies.

Abstract

Recently, Multi-Contrast MR Reconstruction (MCMR) has emerged as a hot research topic that leverages high-quality auxiliary modalities to reconstruct undersampled target modalities of interest. However, existing methods often struggle to generalize across different k-space undersampling patterns, requiring the training of a separate model for each specific pattern, which limits their practical applicability. To address this challenge, we propose UniFS, a Unified Frequency-Spatial Fusion model designed to handle multiple k-space undersampling patterns for MCMR tasks without any need for retraining. UniFS integrates three key modules: a Cross-Modal Frequency Fusion module, an Adaptive Mask-Based Prompt Learning module, and a Dual-Branch Complementary Refinement module. These modules work together to extract domain-invariant features from diverse k-space undersampling patterns while dynamically adapt to their own variations. Another limitation of existing MCMR methods is their tendency to focus solely on spatial information while neglect frequency characteristics, or extract only shallow frequency features, thus failing to fully leverage complementary cross-modal frequency information. To relieve this issue, UniFS introduces an adaptive prompt-guided frequency fusion module for k-space learning, significantly enhancing the model's generalization performance. We evaluate our model on the BraTS and HCP datasets with various k-space undersampling patterns and acceleration factors, including previously unseen patterns, to comprehensively assess UniFS's generalizability. Experimental results across multiple scenarios demonstrate that UniFS achieves state-of-the-art performance. Our code is available at https://github.com/LIKP0/UniFS.

Figures (4)

• Figure 1: Visualization of multi-contrast images. The k-space magnitudes (h) reveal nearly identical high-frequency components (center, structural information) with distinct low-frequency distributions (sides, style information), while the gradient profiles (g) reflect similar global structures with localized variations, highlighting the complementary roles of spatial (gradient) and frequency (k-space) domains in multi-contrast MRI representation.
• Figure 2: Reconstructing T2WIs (target) with T1WIs (reference) from undersampled k-space data under different undersampling patterns (2nd row).
• Figure 3: The overall architecture of UniFS for multi-contrast MRI reconstruction. The network iteratively refines various undersampled k-space data using a frequency-spatial dual-branch integration mechanism. This mechanism incorporates Cross-Modal Frequency Fusion (CMF) with Adaptive Mask-Based Prompt Learning (AMPL) and Dual-Branch Complementary Refinement (DCR).
• Figure 4: Visualization of reconstructed images (1st, 3rd rows) and error maps (2nd, 4th rows) for different methods on BraTs with 4$\times$ acceleration. The zoomed-in regions (highlighted by red boxes) show that our method achieves higher fidelity, as evidenced by the corresponding error maps.