数据与模型双驱动的高效压缩感知磁共振成像重构算法  

An Efficient Data-Model Dual-Drive Algorithm for Compressed Sensing MRI

作　　者：张宇夕 马龙[1,2] 刘日升 程世超[2,3] 樊鑫 罗钟铉 Zhang Yuxi;Ma Long;Liu Risheng;Cheng Shichao;Fan Xin;Luo Zhongxuan(DUT-RU International School of Information Science&Engineering,Dalian University of Technology,Dalian 116620;Key Laboratory for Ubiquitous Network and Service Software of Liaoning Province,Dalian 116620;School of Mathematical Science,Dalian University of Technology,Dalian 116620;Institute of Artificial Intelligence,Guilin University of Electronic Technology,Guilin 541004)

机构地区：[1]大连理工大学-立命馆大学国际信息与软件学院,大连116620 [2]辽宁省泛在网络与服务软件重点实验室,大连116620 [3]大连理工大学数学科学学院,大连116620 [4]桂林电子科技大学人工智能研究所,桂林541004

出　　处：《计算机辅助设计与图形学学报》2020年第6期903-910,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61672125,61733002,61632019,61572096)。

摘　　要：传统压缩感知磁共振成像重构算法基于先验构造与迭代求解,通常具有很低的计算效率,近期提出的深度方法依赖训练数据与结构设计,因此泛化能力差.针对两者的问题,提出一种高效鲁棒的重构算法以实现性能与效率的平衡.算法从互补的视角出发,对细节恢复和伪影去除2个问题分别构建模型驱动的先验表达过程与数据驱动的深度预测过程,实现了领域知识与深度信息的充分融合;交替迭代的求解机制保证中间结果被及时修正,进一步引导解序列沿着理想的传播方向逼近目标解.针对T1加权与T2加权数据的实验结果表明,与现有先进算法相比,所提算法在3种采样模板与5种采样频率下均能实现更高的重构精度,且提高了在GPU与CPU上的计算效率,进一步实验表明所提算法对采样部位差异与莱斯噪声干扰具有更强的鲁棒性.Traditional compressed sensing MRI methods that focus on constructing better prior regularizations or numerical iterative optimizations usually suffer from heavy computational burden.Recently developed deep learning based approaches rely too much on the selection of training data and deep architecture,thus have poor abilities of generalization.To address these issues,we propose an efficient and robust algorithm to achieve the balance between reconstruction accuracy and efficiency.We construct a model-driven priori expression process and a data-driven prediction process for details restoration and artifacts correction,in a complementary perspective,realizing an integration of domain knowledge and deep representation.Further,the iteratively alternating mechanism ensures that the output propagation can be corrected in time and guided towards the desired solution in expected direction.Detailed experiments on T1 and T2 weighted data demonstrate that compared with the state-of-the-art,our method achieves higher reconstruction accuracy for all three kinds of sampling patterns and five sampling ratios,as well as higher computation efficiency on both GPU and CPU.Further experiments show that our method provides stronger robustness to data variations and noise pollution.

关 键 词：压缩感知磁共振成像 深度学习 邻近梯度 残差学习 凸优化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]