A Massively Parallel Approach to Affine Transformataion in Medical Image Registration
A Massively Parallel Approach to Affine Transformataion in Medical Image Registration
复制标题
DOI：
10.1109/hpcc.2009.13
复制
发表时间：
2009-06
期刊：
影响因子：
--
通讯作者：
H. V. Luong;Jong-Myon Kim
中科院分区：
文献类型：
--
作者：
H. V. Luong;Jong-Myon Kim
医学图像配准在研究疾病过程、了解正常发育和衰老过程中起着重要作用。在大多数医生注册方法中，仿射变换是必不可少的组成部分。仿射变换由线性变换(旋转和缩放)和平移的任意组合组成。这些算法通常在计算上是昂贵的。并行计算机的日益普及使得将这些任务并行化成为一个有吸引力的选择。本文提出了一种大规模并行仿射变换的方法，使用一个典型的数据并行体系结构来加速这类算法。我们的并行方法的结果在处理性能和能源效率方面都是突出的。与使用诸如TI、DSP和ARM系列等商业处理器的其他实现相比，所提出的并行方法实现了三阶计算能力和二阶能效。
Medical image registration plays an important role in investigating disease processes and understanding normal development and ageing. An essential component in most medical registration approaches is affine transformation. The affine transformation is made up of any combination of linear transformations (rotation and scaling) followed by translation. These algorithms are generally computationally expensive. The increasing availability of parallel computers makes parallelizing these tasks an attractive option. This paper proposes a massively parallel approach for affine transformations using a representative data parallel architecture to accelerate such algorithms. The result of our parallel approach is outstanding in terms of both processing performance and energy efficiency. The proposed parallel approach achieves a three order of computational capabilities and a second order of energy efficiency of other implementations using commercial processors such as TI DSP and ARM families.