Large-scale Parallel Katsevich Algorithm for 3D Cone-beam CT Image Reconstruction

描述（申请人提供）：医学图像重建需要高性能计算（HPC）和高端计算资源。它们对于CT/微CT医学成像尖端技术的实际实施至关重要。虽然最近开发的算法非常复杂，但它们需要大量的时间来重建三维图像。几十年来，寻找一种经济高效的并行算法和高性能的计算系统一直是一个挑战。爱荷华大学最近开发的用于三维锥束CT图像重建的并行Katsevich算法促使研究人员开发了一种大规模的并行Katsevich算法。该算法将用于利用集成了大量处理器的NSF TeraGrid系统进行高分辨率CT/micro-CT医学图像重建。本课题的总体目标是为大规模异构系统的三维CT/微CT医学图像重建开发一种特定的并行算法。这种并行算法将允许医学研究人员和/或临床专业人员在集成多个高性能计算集群的大规模分布式计算系统上实现高分辨率、3d医学图像重建的高性能。这个R21项目的具体目标是：(1)在高性能计算系统上开发大规模并行Katsevich算法，重点是内存分配、投影数据分解和可扩展性；(2)开发分布式环境下兼顾负载均衡、容错和网络影响的功能；(3)利用TeraGrid超级计算资源，从加速、并行效率、可扩展性、粒度和网络延迟等方面评估并行性能的计算基准。

DESCRIPTION (provided by applicant): Medical image reconstruction requires high performance computing (HPC) and high-end computing resources. They are extemely critical for practical implementation of cutting-edge technology in CT/micro-CT medical imaging. Although the recently developed algorithms are very sophisticated, they require significant time for 3-D image reconstruction. It has been a challenge for decades to find an economic and efficient parallel algorithm, and high performance computing system. The recently developed parallel Katsevich algorithm for 3D cone-beam CT image reconstruction at the University of Iowa has prompted the investigators to develop a large-scale parallel Katsevich algorithm. This algorithm will be developed and implemented for high-resolution CT/micro-CT medical image reconstructions using NSF TeraGrid system which integrates a massive number of processors. The overall goal of this proposal is to develop a specific parallel algorithm for 3-D CT/micro-CT medical image reconstruction on large scale heterogeneous systems. This parallel algorithm will allow medical researchers and/or clinical professionals to achieve high-performance for high-resolution, 3-D medical image reconstruction on a large-scale distributed computing system integrating multiple HPC clusters. The specific aims of this R21 project are to (1) develop large-scale parallel Katsevich algorithm on high performance computing systems with focuses on memory allocation, projection data decomposition, and scalability; (2) develop functions which can account for load balancing, fault-tolerance, and network impact in distributed environment; (3) compute benchmarks for evaluation of parallel performance in terms of speed-up, parallel efficiency, scalability, granularity, and network latency, using TeraGrid supercomputing resources.