作为辐射传输方程的高阶近似, 简化球谐近似模型成为近年光学分子成像研究的重点, 但计算效率低限制了它的广泛应用, 为此提出一种基于图形处理器的并行加速策略, 采用NVIDIA 公司推出的统一计算设备架构,对求解过程中耗时最多的两个模块——有限元刚度矩阵的生成和线性方程组的求解进行基于图形处理器的并行加速; 根据统一计算设备架构的特点, 进行计算任务的分配、存储器的合理使用以及数据的预处理三方面的优化; 仿体及数字鼠仿真实验对比刚度矩阵生成时间以及平均迭代时间, 以评价所提出方法的加速效果。实验结果表明, 该方法可使求解速度提高30倍左右, 展示了该方法在光学分子成像中的优势及潜力。

As a high-order approximation model to Radiative Transfer Equation, simplify spherical harmonic (SPN) approximation has become a hot research topic in optical molecular imaging research. However, low computational efficiency imposes restrictions on its wide applications. This paper presented a graphics processing units (GPU)-parallel accelerated strategy for solving SPN model. The proposed strategy adopted compute unified device architecture (CUDA) parallel processing architecture introduced by NVIDIA Company to build parallel acceleration of two most time-consuming modules, generation of stiffness matrix and solving linear equations. Based on the feature of CUDA, the strategy optimized the parallel computing in tasks distribution, use of memory units and data preprocessing. Simulations on phantom and digital mouse model are designed to evaluate the accelerating effect by comparing the time for system matrix generation and average time of each step iteration. Experimental results show that the overall speedup ratio is around 30 times, which exhibit the advantage and potential of the proposed strategy in optical molecular imaging.