TRD3 - Image Reconstruction

TRD3 - 图像重建

• 批准号: 10549856
• 负责人: Mehmet Akcakaya
• 金额: $ 22.19万
• 依托单位: UNIVERSITY OF MINNESOTA
• 依托单位国家: 美国
• 财政年份: 2019
• 资助国家: 美国
• 起止时间: 2019-02-01 至 2024-01-31
• 项目状态: 已结题
• 来源: https://reporter.nih.gov/project-details/10549856
• 关键词: 3-Dimensional; Acceleration; Address; Algorithms; Anatomy; Brain imaging; Calibration; Cardiac; Clinical; Data; Data Set; Databases; Dependence; Development; Diffusion; Dimensions; Elements; Exhibits; Formulation; Fourier Transform; Human; Image; Joint repair; Joints; Learning; Machine Learning; Magnetic Resonance Imaging; Magnetic Resonance Spectroscopy; Maps; Measurement; Measures; Methods; Modeling; Morphologic artifacts; Motion; Noise; Perfusion; Phase; Physiological; Play; Protocols documentation; Radial; Research; Residual state; Resolution; Role; Rotation; Sampling; Scanning; Scheme; Slice; System; Techniques; Technology; Training; Variant; computerized tools; connectome; convolutional neural network; cost; data space; deep learning; high resolution imaging; image reconstruction; improved; mathematical model; multidimensional data; novel; novel strategies; optical sensor; reconstruction; resilience; respiratory; sensor; simulation; spatiotemporal; temporal measurement; usability

Project Summary/Abstract Image reconstruction from raw measurements is an inverse problem of fundamental importance in MRI. The basic formulation for such reconstructions involve a k-space sampled uniformly on a Cartesian grid at greater than the Nyquist rate, which is Fourier transformed to generate the desired image. However, this acquisition- reconstruction strategy is often difficult to perform in practical research and clinical settings, as it leads to long scan times, necessitating trade-offs in spatial and temporal resolutions. This observation has led to the development of multiple reconstruction strategies over the last few decades, including partial Fourier imaging, parallel imaging, non-Cartesian acquisitions and compressed sensing, where the reconstruction goes beyond a simple Fourier transform, and often involves careful mathematical modeling of the MR system and images. The aforementioned developments aim to address a continuous need for faster imaging, improved resolutions and robustness, both in clinical and research settings. However, as the existing methods reach the limits of resolution and acceleration achievable in the presence of system and physiological limitations, new reconstruction strategies are needed to improve image quality for various acquisition strategies. In this TRD, we seek to develop new image reconstruction techniques for enabling fast high-resolution acquisitions, improving noise resilience, allowing for different encoding strategies, while increasing robustness to underlying physiological and system variations. Our developments for fast high-resolution imaging include improved strategies for k-space interpolation reconstruction in Cartesian imaging, as well as new self- calibrated techniques for three-dimensional non-Cartesian imaging. For the former, we extend the liner shift- invariant convolutional interpolation approaches for reconstructing multi-coil data in two ways: i) Scan-specific deep learning without training databases for non-linear estimation of missing k-space data, in simultaneous multi-slice, parallel and partial Fourier imaging, ii) Region-specific shift-variant linear kernels for highly- accelerated volumetric parallel imaging. For non-Cartesian acquisitions, our self-calibration is used to estimate radius- and rotation-specific interpolation kernels, without additional ACS data. We also tackle the problem of improving non-Fourier encoded acquisitions, such as spatiotemporal encoding, and devise fast matrix sparsifying approaches to enable regularized reconstructions without high computational burden. To further improve reconstruction fidelity in multi-dimensional acquisitions, we propose the local use of high-order tensor models, along with an information theoretic approach for parameter-free regularization. Finally, we consider imaging in the presence of physiological and system variations, such as motion and B0 inhomogeneities, which are especially pronounced at ultrahigh field strengths, and develop a self-consistency based framework for nonlinear inversion, which utilizes improved initialization from external sensors or sequence elements.

项目总结/摘要 从原始测量数据重建图像是磁共振成像中一个非常重要的逆问题。的 这种重建的基本公式涉及在笛卡尔网格上以更大的速度均匀采样的k空间 而不是奈奎斯特速率，后者经过傅立叶变换以生成所需图像。不过，这次收购- 重建策略通常难以在实际研究和临床环境中执行，因为它导致长时间的 扫描时间，需要在空间和时间分辨率上进行权衡。这一观察导致了 在过去的几十年里，多种重建策略的发展，包括部分傅立叶成像， 并行成像，非笛卡尔采集和压缩感知，其中重建超越了 简单的傅立叶变换，并且通常涉及MR系统和图像的仔细数学建模。的 上述发展旨在解决对更快成像、提高分辨率和 在临床和研究环境中的稳健性。然而，由于现有的方法达到了 分辨率和加速度可实现的系统和生理限制，新的 需要重建策略来改善各种采集策略的图像质量。 在这个TRD中，我们寻求开发新的图像重建技术，以实现快速高分辨率 采集，提高抗噪声能力，允许不同的编码策略，同时提高鲁棒性 潜在的生理和系统变化。我们的快速高分辨率成像开发包括 笛卡尔成像中k空间插值重建的改进策略，以及新的自 用于三维非笛卡尔成像的校准技术。对于前者，我们延长了班轮班次- 用于以两种方式重建多线圈数据的不变卷积插值方法： 深度学习，无需训练数据库，用于缺失k空间数据的非线性估计，同时 多切片、并行和部分傅立叶成像，ii）用于高度- 加速体积并行成像。对于非笛卡尔采集，我们的自校准用于估计 特定于半径和旋转的插值内核，无需额外的ACS数据。我们还解决了以下问题： 改进非傅立叶编码采集，如时空编码，并设计快速矩阵 稀疏化方法以使得能够进行正则化重建而没有高计算负担。进一步 为了提高多维采集中的重建保真度，我们提出了高阶张量的局部使用 模型，沿着与信息理论的方法，无参数正则化。最后，我们考虑 在存在生理和系统变化的情况下成像，例如运动和B 0不均匀性， 特别是在磁场强度明显，并制定一个自我一致性为基础的框架， 非线性反演，其利用来自外部传感器或序列元件的改进的初始化。