Department of Biomedical Engineering, Case Western Reserve University, Cleveland, Ohio, USA.
J Magn Reson Imaging. 2010 Feb;31(2):457-65. doi: 10.1002/jmri.22051.
To develop a fast and robust Iterative Decomposition of water and fat with Echo Asymmetry and Least-squares (IDEAL) reconstruction algorithm using graphics processor unit (GPU) computation.
The fat-water reconstruction was expedited by vectorizing the fat-water parameter estimation, which was implemented on a graphics card to evaluate potential speed increases due to data-parallelization. In addition, we vectorized and compared Brent's method with golden section search for the optimization of the unknown field inhomogeneity parameter (psi) in the IDEAL equations. The algorithm was made more robust to fat-water ambiguities using a modified planar extrapolation (MPE) of psi algorithm. As compared to simple planar extrapolation (PE), the use of an averaging filter in MPE made the reconstruction more robust to neighborhoods poorly fit by a two-dimensional plane.
Fat-water reconstruction time was reduced by up to a factor of 11.6 on a GPU as compared to CPU-only reconstruction. The MPE algorithms incorrectly assigned fewer pixels than PE using careful manual correction as a gold standard (0.7% versus 4.5%; P < 10(-4)). Brent's method used fewer iterations than golden section search in the vast majority of pixels (6.8 +/- 1.5 versus 9.6 +/- 1.6 iterations).
Data sets acquired on a high field scanner can be quickly and robustly reconstructed using our algorithm. A GPU implementation results in significant time savings, which will become increasingly important with the trend toward high resolution mouse and human imaging.
利用图形处理器单元（GPU）计算开发一种快速、稳健的迭代分解水脂反演和最小二乘（IDEAL）重建算法。
通过矢量化水脂参数估计来加速水脂重建，在图形卡上实现以评估数据并行化带来的潜在速度提升。此外，我们矢量化了 Brent 方法与黄金分割搜索，并比较了它们在 IDEAL 方程中对未知场不均匀性参数（psi）的优化。通过修改 psi 算法的平面外推（MPE），使算法对水脂模糊性更稳健。与简单的平面外推（PE）相比，MPE 中使用平均滤波器使重建对拟合不理想的二维平面邻域更稳健。
与仅使用 CPU 重建相比，GPU 上的水脂重建时间最多可减少 11.6 倍。使用仔细的手动校正作为金标准，MPE 算法错误分配的像素数比 PE 少（0.7%比 4.5%；P < 10(-4)）。在大多数像素中，Brent 方法的迭代次数都少于黄金分割搜索（6.8 +/- 1.5 比 9.6 +/- 1.6 次迭代）。
使用我们的算法可以快速、稳健地重建高场扫描仪采集的数据集。GPU 实现可显著节省时间，随着高分辨率鼠标和人体成像的趋势，这一点将变得越来越重要。
