Frequency Domain-based Matching and Integration Network for Multi-contrast MRI Reconstruction

Multi-contrast magnetic resonance (MR) images are crucial for diagnosing diseases and analysis in clinical practice, yet their acquisition often entails long scanning procedures. Recently, exploiting shared information among multi-contrast MR images yields favorable results in their reconstruction. However, most studies simply concatenate the multi-contrast information without effective matching and fusion mechanisms to mitigate the impact of redundant features. Furthermore, the feature extraction backbone networks tend to lose high-frequency details like textures during downsampling. To promote the development of this research field, we develop a matching and fusion framework for multi-contrast MR reconstruction and design a feature extractor that explores frequency-domain information. Firstly, our feature extractor captures both long-range and short-range dependencies in the target and reference images, acquiring high- and low-frequency feature information. Then, we introduce a progressive feature matching mechanism to identify features in the reference image that resemble those in the target. Additionally, we propose an innovative fusion strategy that leverages non-local Fourier self-attention and cross-attention mechanisms to precisely combine the semantic feature from multi-contrast MR images. Extensive experiments validate the superior performance of our method in multi-contrast MR image reconstruction, surpassing other state-of-the-art methods.