High performance cluster computing with 3-D nonlinear diffusion filters

Abstract

This paper deals with parallelisation and implementation aspects of PDE-based image processing models for large cluster environments with distributed memory. As an example we focus on nonlinear diffusion filtering which we discretise by means of an additive operator splitting (AOS). We start by decomposing the algorithm into small modules that shall be parallelised separately. For this purpose image partitioning strategies are discussed and their impact on the communication pattern and volume is analysed. Based on the results we develop an algorithmic implementation with excellent scaling properties on massively connected low latency networks. Test runs on a high-end Myrinet cluster yield almost linear speedup factors up to 209 for 256 processors. This results in typical denoising times of 0.5 seconds for five iterations on a 256 × 256 × 128 data cube.

FPGA-based real-time 3D image preprocessing for image-guided medical interventions

“…The work reported in this manuscript introduces a novel FPGA-based implementation of 3D anisotropic diffusion filtering. The reported implementation, using a single FPGA device, achieves a voxel processing rate better than that previously achieved by a 256-processor cluster [5]. The reported solution is compact, offers real-time performance and hence is more suitable for clinical deployment.…”

“…Earlier efforts to accelerate 2D anisotropic diffusion filtering employed graphics hardware [34] and analog hardware [17,38]. Accelerated implementations of 3D realizations using computing clusters have also been reported [4,5,35]. The work reported in this manuscript introduces a novel FPGA-based implementation of 3D anisotropic diffusion filtering.…”

“…(5). As indicated by that formulation, we use a Gaussian-filtered version of the image for improved diffusion coefficient estimation.…”

“…Earlier attempts to accelerate 3D anisotropic diffusion filtering were targeted toward multiprocessor clusters [5,35]. Despite the near-linear speedup offered by these techniques, the need to employ up to 256 processors to achieve real-time performance makes them less suitable for clinical deployment.…”

“…Some researchers have recently reported high-speed implementations of 3D median filtering using graphics processing units [36] and FPGAs [23]. This manuscript presents an FPGA-based 3D median filtering module that is faster than currently existing solutions and supports higher 3D kernel sizes (3,5,7). The reported architecture can achieve a processing rate close to 200 Megavoxels per second for both the 3D anisotropic diffusion and 3D median filtering, which is equivalent to about 50 processing iterations or operations per second for images of size 256 · 256 · 64.…”