PFONet:一种用于加速MRI的渐进式聚焦导向双域重建网络
PFONet:A progressive focus-oriented dual-domain reconstruction network for accelerated MRI
王钟贤 1王志文 1张中洲 1杨子元 1冉茂松 1余慧 1张意2
作者信息
- 1. 四川大学计算机学院,成都 610065
- 2. 四川大学计算机学院,成都 610065;四川大学网络空间安全学院,成都 610065
- 折叠
摘要
对k空间数据进行欠采样操作是一种有效的加速磁共振(MRI)成像方法,但是利用欠采样的数据准确地进行图像重建是一个具有挑战性的工作.最近,利用频域和图像域双域信息的神经网络重建方法可以提升MRI重建性能,因此引起了研究者的关注.然而,这些方法主要存在以下两个不足:首先,在频率域,这些带有传统归一化模块的方法将测量数据和零填充区域同等对待,导致了k空间特征偏移现象和次优的重建效果;其次,在图像域,现有方法通常忽略了多尺度的动态特征对于细节恢复的重要性,因此传统图像域网络难以学习足够的全局-局部信息以重建结构细节.因此,本文提出了一种新型渐进式聚焦导向双域重建网络(PFONet),以分别克服频域和图像域的这些限制.在频域,提出了一个专注于零填充区域的区域归一化模块,逐步缓解特征偏移问题,并预测可靠的k空间数据.在图像域,提出了一个带有通道级门控机制的动态注意力模块,专注于从多尺度感受野中提取丰富的全局-局部特征,以恢复细节.定量和定性实验表明,在与几种最先进方法的比较中,本文提出的PFONet最为轻量,同时实现了更优的重建性能.
Abstract
Undersampling k-space data to accelerate Magnetic Resonance Imaging(MRI)is effective but challenging for accurate image reconstruction.Recently,neural networks,particularly dual-domain models that leverage both frequency and image domain data,have gained attention for enhancing MRI reconstruction.However,these methods are inadequate for two main reasons.Firstly,in the frequency domain,these meth-ods with traditional normalization modules treat measurements and zero-filled areas equally,leading to feature shift and suboptimal reconstruction.Secondly,in the image domain,existing methods typically ignore multi-scale features and lack dynamic prosperity,thus being challenging for networks to learn sufficient global-local information to preserve structural details.Here,the authors present a novel progressive focusoriented dual-domain reconstruction network(PFONet)to overcome these limitations in frequency and image domains,re-spectively.In the frequency domain,the authors propose an area normalization module that focuses on the zero-filled areas,progressively mitigates the feature shift,and predicts reliable k-space data.In the image do-main,the authors propose a dynamic attention module with a channel-wise gating mechanism to focus on rich global-local feature extraction from multi-scale receptive fields for detail recovery.Quantitative and qualitative experiments demonstrate that our PFONet achieves competitive performance compared to several state-of-the-art methods while the proposed network is lightweight.
关键词
磁共振成像/磁共振重建/渐进式聚焦导向/多尺度特征聚合Key words
Magnetic Resonance Imaging(MRI)/MRI reconstruction/Progressive focus-oriented/Multi-scale feature aggregation引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据