A Compute Cluster for Brain Imaging and Analysis

 描述：使用高分辨率功能和结构磁共振成像(包括扩散张量成像，或DTI)的神经成像研究涉及采集和分析非常大的数据集，然后必须通过一系列计算密集型处理任务将这些数据集转换为有意义的信息。该项目名为“用于脑成像和分析的计算机群”，由杜克大学NIH支持的研究人员提交，他们主要在杜克大学医学院的杜克-北卡罗来纳大学脑成像和分析中心进行神经成像研究。合作的科学家们发现，改进的采集和分析方法带来的海量数据，使得及时全面分析所有实验室的数据变得越来越困难。在我们现有的计算基础设施下，一些过程，如通过大规模并行成像(面内和多波段直通平面)进行超高分辨率(亚毫米)扩散磁共振成像，对于典型的10次全脑功能检查将需要20多个小时。此外，用于我们的超高分辨率磁化率映射的特定大型矩阵求逆将需要每个节点高达700 GB的内存。因此，我们正在申请资金购买高性能计算集群，以满足这些和其他处理需求。我们的建议基于我们过去管理60个节点的Linux群集(Dell PowerEdge 1950)的经验，该群集具有高速光纤通道适配器并连接到我们的多个数据服务器。我们预计，这种大大改进的计算资源将极大地增强我们的能力，以满足NIH赞助的许多研究项目中大幅增加的MRI采集能力和分析需求，并以更低的总体成本获得最大效率。除了满足我们神经成像科学家的计算需求外，这一提议还受益于我们之间已经存在的密切合作，在杜克大学的大数据倡议中产生了更广泛的影响。

 DESCRIPTION: Neuroimaging research using high-resolution functional and structural MRI (including diffusion tensor imaging, or DTI) involves the acquisition and analysis of very large data sets that then must be transformed through a series of computation-intensive processing tasks into meaningful information. This project, entitled "A Compute Cluster for Brain Imaging and Analysis", is submitted by NIH-supported investigators at Duke University who primarily conduct neuroimaging research at the Duke - UNC Brain Imaging and Analysis Center at Duke University Medical School. The collaborating scientists have found that the sheer volume of data resulting from improved acquisition and analysis methods has made it increasingly difficult to fully analyze data from all the labs in a timely manner. With our existing computational infrastructure, some processes such as the ultrahigh resolution (submillimeter) diffusion MRI through massive parallel imaging (both in-plane and multi-band through-plane) would take over 20 hrs for a typical ten-run whole-brain functional exam. Also, specific large matrix inversions fo our ultrahigh resolution susceptibility mapping would require memory of up to 700 GB per node. Thus, we are requesting funds to purchase a high- performance compute cluster to meet these and other processing requirements. Our proposal builds upon our past experience administering a 60-node, though quite dated, Linux cluster (Dell PowerEdge 1950), with high-speed fiber channel adaptors and connections to our multiple data servers. We anticipate that this much improved computational resource will greatly enhance our ability to meet the much increased MRI acquisition capability and analysis demand in the many NIH-sponsored research projects, and gain maximum efficiency at lower overall cost. In addition to meeting the computational needs of our neuroimaging scientists, this proposal also benefits from the close collaboration that already exists among us in Duke's big data initiative to have a wider impact.