GPU implementation of parallelized microwave tomography algorithm

Abstract

Microwave tomography (MWT) has good potential to be used for medical imaging, however, most of MWT algorithms rely on local optimization methods and need regularization to find the a solution to inverse scattering problems. By using global optimization method for optimization, the non-deterministic nature of the genetic algorithm allows the inverse solver to avoid local minima without the use of regularization methods such as Tikhonov regularization. By not relying on regularization assumptions, high contrast areas of the imaging target can be resolved, whereas regularizations assume smooth dielectric contrast gradients. Resolving areas of high permittivity contrast is necessary to detect small tumors, less than a millimeter in length, as required for effective treatment. Our goal is to implement a fast MWT algorithm based on Finite Difference Time Domain (FDTD) forward solver and global optimization methods. In this regards, we propose the use of graphics processing unit (GPU) for FDTD computation. We have developed a FDTD program using NVidia's CUDA C language. The GPU implemented FDTD simulation was tested to yield 100-fold speed increase from standard Central Processing Unit (CPU) FDTD simulations.