Compound Attention and Neighbor Matching Network for Multi-contrast MRI Super-resolution

作     者：Chen, Wenxuan Wu, Sirui Wang, Shuai Li, Zhongsen Yang, Jia Yao, Huifeng Song, Xiaolei 

作者机构：The Center for Biomedical Imaging Research Tsinghua University Beijing100084 China The School of Materials Science and Engineering Tsinghua University Beijing100084 China The Department of Computer Science and Engineering Hongkong University of Science and Technology Hong Kong 

出 版 物：《arXiv》 (arXiv)

年 卷 期：2023年

核心收录：

主　　题：Magnetic resonance imaging 

摘      要：Multi-contrast magnetic resonance imaging (MRI) reflects information about human tissue from different perspectives and has many clinical applications. Multi-contrast super-resolution (SR) of MRI could synthesize high-resolution images of one modality from the acquired low-resolution (LR) images, by utilizing the complementary information from another modality. An important prior knowledge of multi-contrast MRI is that, multi-contrast images are obtained with similar or same field-of-views (FOVs). However, existing methods did not exploit this prior;only performing simple concatenation of the reference and LR features, or global feature-matching. Herein, we proposed a novel network architecture with compound-attention and neighbor matching (CANM-Net) for multi-contrast MRI SR. Specifically, to effectively utilize the priors of similar FOVs, CANM-Net proposed a neighborhood-based feature-matching method, which only calculates the similarity of a LR patch with the corresponding patch on the HR reference and its adjacent ones;the compound self-attention with a pyramid-structure effectively captures the dependencies in both spatial and channel dimension. We conduct experiments on the IXI, fastMRI, and in-house datasets, with T2-, T1-weighted images as the LR and reference images, respectively. In 4× and 2× SR tasks, the CANM-Net outperforms state-of-the-art approaches in both retrospective and prospective experiments. The ablation study proves the rationality of CANM-net. Additionally, when the input LR images and the reference HR are imperfectly registered, CANM-Net still achieves the best performance among all test methods. In summary, CANM-Net may have potential in clinical applications, for recovery and enhancement of multi-contrast MRI. Copyright © 2023, The Authors. All rights reserved.