High performance single and multi-GPU acceleration for Diffuse Optical Tomography

Abstract

Diffuse Optical Tomography (DOT) is a diagnostic imaging modality, where optical parameters such as absorption and scattering coefficient distributions inside the living tissue are recovered to understand the structural and functional variations in the tissue under study. The numerical method of DOT image reconstruction is an iterative process that demands high computational power, especially in the case of recovering fully three dimensional (3D) optical property distribution inside a complex geometry such as human head which hampers physician to view reconstructed images and monitor a patient in real time. In order to reconstruct 3D DOT images at a high speed, Broyden method based iterative image reconstruction algorithm and a parallelization strategy are employed in CUDA parallel computing platform to utilize tremendous computational power of GPU. Three different single GPU systems equipped with Nvidia Tesla C2070, Tesla k20c and Tesla k40 respectively, and a muti-GPU (two Tesla M2090 GPUs) in a computing node in a HPC cluster are used to evaluate computation performance due to algorithmic improvement and GPU parallel computation. We have used three dimensional finite element method (FEM) and discretized an infant head into 45702 tetrahedral elements and 8703 nodes to solve the forward and inverse problems. We have achieved a significant speedup for the 3D DOT image reconstruction of the head phantom.