Pattern Recognition Laboratory, Universität Erlangen-Nürnberg, 91058 Erlangen, Germany.
Med Phys. 2011 Jan;38(1):468-73. doi: 10.1118/1.3525838.
Interventional reconstruction of 3-D volumetric data from C-arm CT projections is a computationally demanding task. Hardware optimization is not an option but mandatory for interventional image processing and, in particular, for image reconstruction due to the high demands on performance. Several groups have published fast analytical 3-D reconstruction on highly parallel hardware such as GPUs to mitigate this issue. The authors show that the performance of modern CPU-based systems is in the same order as current GPUs for static 3-D reconstruction and outperforms them for a recent motion compensated (3-D+time) image reconstruction algorithm.
This work investigates two algorithms: Static 3-D reconstruction as well as a recent motion compensated algorithm. The evaluation was performed using a standardized reconstruction benchmark, RABBITCT, to get comparable results and two additional clinical data sets.
The authors demonstrate for a parametric B-spline motion estimation scheme that the derivative computation, which requires many write operations to memory, performs poorly on the GPU and can highly benefit from modern CPU architectures with large caches. Moreover, on a 32-core Intel Xeon server system, the authors achieve linear scaling with the number of cores used and reconstruction times almost in the same range as current GPUs.
Algorithmic innovations in the field of motion compensated image reconstruction may lead to a shift back to CPUs in the future. For analytical 3-D reconstruction, the authors show that the gap between GPUs and CPUs became smaller. It can be performed in less than 20 s (on-the-fly) using a 32-core server.
从 C 臂 CT 投影重建三维容积数据是一项计算密集型任务。硬件优化不是可选的，而是强制性的，因为它对性能要求很高，这对于介入图像处理，特别是图像重建至关重要。由于对性能的高要求，已经有几个小组在 GPU 等高度并行的硬件上发布了快速分析三维重建，以解决这个问题。作者表明，对于静态三维重建，现代基于 CPU 的系统的性能与当前的 GPU 相当，并且对于最近的运动补偿（三维+时间）图像重建算法，其性能优于 GPU。
这项工作研究了两种算法：静态三维重建和最近的运动补偿算法。使用标准化的重建基准 RABBITCT 进行评估，以获得可比的结果和另外两个临床数据集。
作者展示了一种参数 B 样条运动估计方案，其中导数计算需要对内存进行多次写操作，在 GPU 上性能不佳，可以从具有大型缓存的现代 CPU 架构中受益。此外，在一个 32 核英特尔至强服务器系统上，作者实现了使用的核心数量的线性扩展，并且重建时间几乎与当前的 GPU 相同。
运动补偿图像重建领域的算法创新可能会导致未来重新转向 CPU。对于分析三维重建，作者表明 GPU 和 CPU 之间的差距已经缩小。使用 32 核服务器，它可以在不到 20 秒（实时）内完成。
