Accelerating advanced MRI reconstructions on GPUs

描述（由申请人提供）：本提案的主要假设是，由于缺乏可访问的计算资源，临床磁共振成像（MRI）目前在空间分辨率、扫描时间和信噪比的权衡方面受到限制，无法实现高级MRI采集和重建方法的临床应用。虽然在研究中使用了先进的MRI采集和重建技术，但临床实用性要求在图像采集的数量级（一分钟或几分钟）内完成图像重建。该提案将开发、验证和基准测试一个灵活的软件包，以允许在基于GPU的商品个人计算机和集群提供的已经广泛、经济和计算效率高的众核计算平台上执行高级MRI重建。具体而言，将创建一个基于GPU的图像重建框架，该框架具有与C代码和Matlab的简单接口，允许用户对3D非笛卡尔轨迹采集的数据进行重建;利用多个接收器线圈进行并行成像;补偿与长数据采集读数相关的磁场不均匀性;并将先前的解剖信息纳入图像重建。这些技术将通过模拟、体模和人体MRI采集进行验证，指标包括计算时间、归一化均方根误差和噪声方差。该软件将与自动优化例程打包，以实现在各种计算平台上的快速执行，包括PC和集群中的多核CPU和众核GPU。该软件，沿着示例重建，样本数据，用户手册和编程文档将通过网络分发，根据开源许可证免费提供给教育用户。在项目结束时，学术和医疗机构的医学物理学家将能够为他们的特定MR采集定制软件，并轻松利用多核CPU和众核GPU计算能力。将所提出的计算实用程序集成到临床中将使当前先进的图像重建技术能够转化为临床，并使下一代MRI诊断技术的开发成为可能。 公共卫生相关性：将开发一个高级图像重建软件库，使临床磁共振成像（MRI）能够利用PC和基于GPU的集群中的多核和众核计算实用程序提供的新兴计算能力。先进的图像重建软件将使临床医学物理学家能够轻松地将自定义成像协议集成到通用MR重建框架中，并将计算速度提高10到100倍。利用这种计算能力，临床成像将能够采用先进的MR采集策略，这将导致比传统MRI采集更短的扫描会话、更高的信噪比和更高的空间分辨率。

DESCRIPTION (provided by applicant): The main hypothesis of this proposal is that clinical magnetic resonance imaging (MRI) is currently limited in its tradeoffs of spatial resolution, scan time, and signal-to-noise by a lack of accessible computational resources to enable clinical application of advanced MRI acquisition and reconstruction methods. While advanced MRI acquisition and reconstruction techniques are used in research, clinical utility requires that image reconstructions be completed in times that are on the order of the image acquisition (one or a few minutes). This proposal will develop, validate, and benchmark a flexible software package to allow for advanced MRI reconstructions to be executed on the already widespread, economical, and computationally-efficient many-core computing platforms offered by GPU-based commodity personal computers and clusters. Specifically, a GPU-based image reconstruction framework will be created with an easy interface to C-code and Matlab that allows users to perform reconstruction of data acquired with 3D non-Cartesian trajectories; utilizing multiple receiver coils for parallel imaging; compensating for magnetic field inhomogeneities associated with long data acquisition readouts; and incorporating prior anatomical information into the image reconstruction. The techniques will be validated through simulation, phantom, and human MRI acquisitions with metrics including computation time, normalized root mean square error, and noise variance. The software will be packaged with automatic optimization routines to enable fast execution on a variety of computational platforms, including both multi-core CPUs and many-core GPUs in PCs and clusters. The software, along with example reconstructions, sample data, user manuals, and programming documents will be distributed through the web, free of charge to educational users in accordance with the open source license. At the conclusion of the project, medical physicists at academic and medical institutions will be able to customize the software for their specific MR acquisitions and easily harness multi-core CPU and many-core GPU computational power. Integration of the proposed computational utility into the clinic will enable translation of current advanced image reconstruction techniques to the clinic and enable development of the next generation of MRI diagnostic technology. PUBLIC HEALTH RELEVANCE: An advanced image reconstruction software library will be developed that allows clinical magnetic resonance imaging (MRI) to harness the emerging computational power provided by multi-core and many-core computational utilities in PCs and GPU-based clusters. The advanced image reconstruction software will allow medical physicists in the clinic to easily integrate custom imaging protocols into the general MR reconstruction framework and reap computational speed-ups on the order of 10 to 100 times. Leveraging this computational power, clinical imaging will be able to adopt advanced MR acquisition strategies that will lead to shorter scan sessions, higher signal-to-noise ratios, and higher spatial resolution than is possible with traditional MRI acquisitions.