基于字典学习理论的快速高分辨率磁共振成像关键技术研究

Since the emergence of the magnetic resonance imaging (MRI), an increasing number of research efforts have been devoted to the acceleration of the imaging process so that a reduced scanning duration and reconstructed images with high resolution can be obtained. Two of the key factors leading to these goals could be the exploitation of the sparsity inherent in the signal/image, and a fast and stable algorithm implementation. In recent years, the gradually popularized theory of sparse and redundant representations has enlightened the signal processing community with novel alternative approaches with improved efficacy. The applicants of this project have proposed a series of augmented Lagrangian-based dictionary learning algorithms, which have been recognized by world-class journals and research societies. To achieve an effective and efficient MR imaging, we intend to carry out further research in the following three main directions: (1) employing the augmented Lagrangian-based decomposition strategy for faster implementations of the proposed algorithms; (2) exploring further improved ability in representation via local geometric information and gradient features during the formulation of the dictionary; (3) incorporating the prior information and constraints of the observed image data into the dictionary learning algorithms for an enhanced capability of parallel imaging. The above three research directions could mutually help obtain improved representations, effectively reduce the amount of data required from scanning, and turn into reality a faster and thus less-invasive imaging process. Conducting this project will concretely achieve the goal of high-resolution imaging with substantially reduced data, promote the applications of dynamic MR imaging and parallel imaging, considerably expand the application ranges of MR imaging in clinical settings.

自磁共振(MRI)系统出现以来，人们一直致力于成像速度的提高，以期既减少扫描时间而又能得到高分辨率的重建图像。如何挖掘信号/图像的稀疏性以及设计快速稳健的实现算法是其成功的两大关键因素。近年来兴起的稀疏表示理论为此开辟了一个有效的途径，申请人在理论上提出了基于增广拉格朗日的字典学习系列算法，积累了一定的成果。为实现高效快速的MRI成像，项目组拟从三个方向为研究切入点：(1)探讨增广拉格朗日分解策略促进算法快速实现；(2)在字典构建过程中结合图像局部几何信息或者图像梯度等信息提高算法模型表示能力；(3)将观测数据约束信息融入于字典学习算法提高并行成像能力。这三个研究角度相辅相成、共同促使字典在构建过程中得到更为丰富的表示结果使扫描数据减少，实现快速成像。本项目的顺利实施将得到大大减少采集数据下的高分辨率图像，推动磁共振动态成像和并行成像等技术的应用，从而大大拓展MRI成像在临床应用的范围。

自磁共振(MRI)系统出现以来，人们一直致力于成像速度的提高，以期既减少扫描时间而又能得到高分辨率的重建图像。如何挖掘信号/图像的稀疏性以及设计快速稳健的实现算法是其成功的两大关键因素。近年来兴起的稀疏表示理论为此开辟了一个有效途径，项目组在增广拉格朗日字典学习系列算法的基础上，从三个方向为切入点展开研究：(1)字典学习基本模型的算法优化实现问题；(2)在字典构建过程中结合图像局部几何信息或者图像梯度等信息提高算法模型表示能力；(3)字典学习算法在磁共振动态成像和并行成像等技术中的应用。项目组集中在稀疏表示及字典优化、梯度域上的处理和复杂成像场景下模型优化三个方面取得系列成果。先后发表论文20余篇，其中在IEEE TMI上和MRM等成像国际权威期刊上发表5篇以上；在ISMRM和ICIP等学术会议专题报告或讲座5人次（Oral presentation 2次以上）。授权快速成像相关发明专利3项。授权软件著作权2项。