基于双域交互Transformer的磁共振图像重建

Dual-domain interaction Transformer for MRI reconstruction

对k空间数据部分采样是加速磁共振成像的主要方法.从欠采样的数据中重建出高质量的磁共振图像,在临床诊断和研究分析中有着重要的应用价值.近年来,基于深度学习的方法在磁共振重建领域取得了一些进展,然而单独面向图像域或频域的网络不能同时利用双域特征共同提升重建质量.另外,虽然已有一些双域重建方法模型,但是缺乏双域数据交互融合限制了重建性能.针对以上问题,本文提出了一种基于双域交互Transformer的磁共振图像重建网络模型,使用Transformer提取双域特征,并利用交互注意力引导双域特征融合,实现了双域特征的高效提取和交互.具体地,首先,由于频域数据每个点都对应着图像域的所有像素点,因此在k空间使用1×1的卷积提取全局特征,同时使用基于窗口的Transformer将注意力的计算限制在了窗口中,减小了计算负担,并且可以有效地对图像域特征进行表示.其次,提出了基于交互注意力的Transformer融合模块引导双域特征融合,通过挖掘双域特征的相关性,实现跨域信息融合.实验证明,在公开的数据集上,本方法相较于其他基线重建方法均能取得更为优异的重建效果.同时,消融实验证明了本文所提出的网络模块的有效性.

Partial k-space sampling is a primary method for accelerating Magnetic Resonance Imaging(MRI).Reconstruction of high-quality MRI images from undersampled data has important applications in clinical diagnosis and research analysis.In recent years,deep learning-based approachs have made some progress in MRI reconstruction.However,the networks that solely focus on the image domain or the frequency domain cannot effectively utilize the features of both domains to improve reconstruction quality.Furthermore,existing dual-domain reconstruction methods often lack interaction and fusion between the domains,limiting their reconstruction performance.To address these issues,a dual-domain interaction Transformer network is proposed in this paper for MRI reconstruction.The proposed method extracts dual-domain features with Transformer and utilizes cross-domain attention to guide the features fusion,enabling efficient extraction and integration of the complementary information from both domains.Specifically,since each point in the frequency domain data corresponds to all pixels in image domain,the 1×1 convolution in k-space is utilized to extract global features,and while a window-based Transformer is designed to model local features in the im‐age domain.Then a fusion module based on cross attention is introduced to guide the fusion of dual-domain features and integrate cross-domain complementary information.Experimental results demonstrate that the method in this paper outperforms other baseline reconstruction methods on publicly available datasets.Moreover,ablation studies validate the effectiveness of the proposed network modules.