high performance computing on graphics processing units: hgpu.org

Marching Cubes (MC) is an algorithm that extracts surfaces from volumetric data. It is used extensively in visualization and analysis of medical data from modalities like CT and MR, often after a 3D segmentation of the interesting structures is performed. Traditional implementations of MC on modern CPUs are slow, using several seconds (even minutes) to return the surface representation before sending it to the Graphics Processing Unit (GPU) for rendering. Fast surface extraction implementations are very beneficial in medical applications where large datasets are processed and time is crucial. Analysis of medical image data usually means changing different parameters so near real-time implementations are very desirable. MC is a completely data-parallel algorithm which makes it ideal for execution on GPUs allowing the result to be rendered on screens in a few milliseconds. In this paper, a MC implementation written in OpenCL that runs entirely on the GPU is presented. We show that our implementation uses a more efficient storage scheme than previous GPU implementations, and that this enables the real-time processing of large medical datasets. Our implementation also shows that GPU implementations written in OpenCL has the potential of being just as fast and efficient as CUDA or shader implementations.