DUAL-DOMAIN CROSS-ITERATION SQUEEZE-EXCITATION NETWORK FOR SPARSE RECONSTRUCTION OF BRAIN MRI Xiongchao Chen12 Yoshihisa Shinagawa1 Zhigang Peng1 Gerardo Hermosillo Valadez1

2025-05-03 0 0 1.67MB 4 页 10玖币

侵权投诉

DUAL-DOMAIN CROSS-ITERATION SQUEEZE-EXCITATION NETWORK FOR SPARSE

RECONSTRUCTION OF BRAIN MRI

Xiongchao Chen1,2∗, Yoshihisa Shinagawa1, Zhigang Peng1, Gerardo Hermosillo Valadez1

1Siemens Healthineers, Malvern, PA 19355, USA

2Department of Biomedical Engineering, Yale University, New Haven, CT 06511, USA

ABSTRACT

Magnetic resonance imaging (MRI) is one of the most com-

monly applied tests in neurology and neurosurgery. However,

the utility of MRI is largely limited by its long acquisi-

tion time, which might induce many problems including

patient discomfort and motion artifacts. Acquiring fewer

k-space sampling is a potential solution to reducing the to-

tal scanning time. However, it can lead to severe aliasing

reconstruction artifacts and thus affect the clinical diagno-

sis. Nowadays, deep learning has provided new insights into

the sparse reconstruction of MRI. In this paper, we present

a new approach to this problem that iteratively fuses the

information of k-space and MRI images using novel dual

Squeeze-Excitation Networks and Cross-Iteration Residual

Connections. This study included 720 clinical multi-coil

brain MRI cases adopted from the open-source deidentiﬁed

fastMRI Dataset [1]. 8-folder downsampling rate was ap-

plied to generate the sparse k-space. Results showed that

the average reconstruction error over 120 testing cases by

our proposed method was 2.28 ±0.57%, which outper-

formed the existing image-domain prediction (6.03 ±1.31%,

p < 0.001), k-space synthesis (6.12 ±1.66%,p < 0.001),

and dual-domain feature fusion (4.05 ±0.88%,p < 0.001).

Index Terms—Dual-domain deep learning, cross-iteration

residual connection, squeeze-excitation network, sparse MRI

reconstruction, multi-coil parallel imaging

1. INTRODUCTION

Magnetic resonance imaging (MRI) has rapidly become an

essential clinical diagnosis and management tool of neurol-

ogy [2]. MRI can discriminate different anatomical structures

with extreme high resolution and contrast, which makes MRI

essential for clinical diagnosis of a wide variety of disorders

including neurological and oncological diseases.

However, the long scanning time of MRI might induce

many problems including patient discomfort, high exam cost,

motion artifacts, and low patient throughput. Thus, reduc-

ing the MRI scanning time is necessary for accurate and ef-

∗Please contact the authors at xiongchao.chen@yale.edu

ﬁcient MRI diagnosis. One approach commonly used in cur-

rent clinical practice for MRI acceleration is Parallel Imaging

[3]. It utilizes multiple receiver coils to simultaneously ac-

quire the multi-coil information. The other potential approach

is downsampling the k-space measurements. However, the

reconstructed images from the downsampled sparse k-space

will display severe aliasing artifacts as shown in Fig. 1, which

largely reduce the accuracy of clinical diagnosis.

Fully-Sampled Sparse Fully-Sampled Sparse

Fig. 1: Fully-sampled and sparse k-spaces and images.

Deep learning has shown promising results in sparse re-

construction of MRI. Existing deep learning methods can

be generally classiﬁed into three categories. The ﬁrst cat-

egory applied the sparsely reconstructed MRI images as

input of neural networks to predict the synthetic fully re-

constructed images [4]. The second category utilized the

sparse k-space as input of networks to generate the syn-

thetic full-view k-space [5]. The third category combines

the features of k-space and images in a dual-domain man-

ner by Fourier Transform, to restore the full-view k-space

[6]. However, the cross-iteration features were ignored in

previous dual-domain methods. In order to incorporate the

cross-iteration information, we present a novel Dual-Domain

Cross-Iteration Squeeze-Excitation Network (DD-CSENet).

Our contribution are: 1) we develop and incorporate the dual

Squeeze-Excitation Network (SENet) into the dual-domain

framework to improve the reconstruction accuracy; 2) we

propose the Cross-Iteration Residual Connection structure

to fuse the information across different iterations to further

improve the network performance.

2. MATERIALS AND METHODS

2.1. Dataset preprocessing

In this study, 720 clinical 16-coil T2-weighted brain MRI

slices were adopted from the fastMRI Dataset [1]. The down-

sampled sparse k-space was generated by masking the fully

arXiv:2210.02523v2 [cs.CV] 13 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

DUAL-DOMAINCROSS-ITERATIONSQUEEZE-EXCITATIONNETWORKFORSPARSERECONSTRUCTIONOFBRAINMRIXiongchaoChen1;2,YoshihisaShinagawa1,ZhigangPeng1,GerardoHermosilloValadez11SiemensHealthineers,Malvern,PA19355,USA2DepartmentofBiomedicalEngineering,YaleUniversity,NewHaven,CT06511,USAABSTRACTMagneticresonanceimagi...

展开>> 收起<<

DUAL-DOMAIN CROSS-ITERATION SQUEEZE-EXCITATION NETWORK FOR SPARSE RECONSTRUCTION OF BRAIN MRI Xiongchao Chen12 Yoshihisa Shinagawa1 Zhigang Peng1 Gerardo Hermosillo Valadez1.pdf

共4页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

DUAL-DOMAIN CROSS-ITERATION SQUEEZE-EXCITATION NETWORK FOR SPARSE RECONSTRUCTION OF BRAIN MRI Xiongchao Chen12 Yoshihisa Shinagawa1 Zhigang Peng1 Gerardo Hermosillo Valadez1

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: