Studies of PC cluster-based parallel processing for large-scale medical images on navigation system of the next generation surgery

以下两个问题在医学图像处理中非常重要。本研究开发了基于PC机群的并行计算程序，该PC机群有64台PC机，每台PC机有两个Pentium-3（1GHz）处理器，2GB内存，连接Myrinet 2000通信网络。(1)体绘制：为了实时绘制至少512 × 512 × 512的体素体，我们开发了一种基于分布式存储多处理器的sort-last并行体绘制方法。我们的sort-last方法由两种方法组成，Hsu的分段光线投射和我们的分屏分层（DSH）合成，其中每个处理器产生一个子图像，并将所有产生的子图像合并到最终图像中。本文介绍了DSH方法，其目的是在大量的处理器上实现高性能的合成：Ourimplementation在PC集群上可以合成一个512 × 512像素的图像约两倍的速度作为现有的方法，二进制交换方法，因此，可以渲染一个512 × 512 × 224体素体积约8帧每秒（fps）。(2)无刚性图像配准：我们的算法通过采用三种技术实现了高分辨率三维（3-D）图像的可扩展配准：（1）数据分布;（2）数据并行处理;（3）动态负载平衡。实验结果表明，我们在PC集群上的并行实现注册肝脏CT图像的512 × 512 × 159体素在8分钟内，而顺序执行需要12小时。此外，我们的实现允许处理器使用更少的存储器，从而使我们能够对齐1024 X 1024 X 590体素图像，这对于单处理器系统是不容易的，由于对存储器空间和处理时间的限制。

The following two problems are very important in medical image processing. In this study, PC cluster-based parallel programs for these problems are developed, This PC cluster has 64 PCs, each of which has two Pentium-3 (1GHz) processors with 2GB main memory, connected with Myrinet 2000 communication network.(1) Volume rendering:To render at least 512 x 512 x 512 voxel volumes in real-time, we have developed a sort-last parallel volume rendering method fqr distributed memory multiprocessors. Our sort-last method consists of two methods, Hsu' s segmented ray casting and our divided-screenwise hierarchical (DSH) composition, in which each processor produces a subimage and merges all the produced subimages into the final image. This paper describes the DSH method, which aims at achieving high performance composition on a large number of processors: Ourimplementation on the PC cluster can composite a 512 x 512 pixel image about twice as fast as an existing method, the binary-swap method, so that can render a 512 x 512 x 224 voxel volume at approximately eight frames per second (fps).(2) No rigid image registration:Our algorithm realizes scalable registration for high-resolution three-dimensional (3-D) images by employing three techniques: (1) data distribution; (2) data-parallel processing; and (3) dynamic load balancing. The experimental results show that our parallel implementation on the PC cluster registers liver CT images of 512 X 512 X 159 voxels within 8 minutes while a sequential implementation takes 12 hours. Furthermore, our implementation allows processors to use less memory, and thereby enables us to align 1024 X 1024 X 590 voxel images, which is not easy for single processor systems due to the restrictions on the memory space and the processing time.