Advisor: Jiri Jaros (Brno University of Technology)
Abstract: Many emerging non-invasive medical procedures such as high intensity focused ultrasound treatment (HIFU), photoacoustic imaging, or transcranial ultrasound therapy require ultrasound wave propagation simulations. Typically soft tissue can be modeled as fluid, weakly heterogeneous medium with frequency dependent absorption and non-linear effects at high ultrasound intensities. The k-Wave acoustic toolbox, widely used in medical community, uses highly efficient k-space pseudo-spectral time-domain (KSTD) discretization of acoustic equations to solve these problems. However, it is difficult to efficiently map the KSTD method onto modern cluster architectures with accelerators such as GPUs. The present thesis analyzes shortcomings of the KSTD method in respect to modern clusters and proposes local Fourier basis approach to improve scaling of the method. It is shown that the proposed method is able to achieve 5x speedup, while having sufficient accuracy for these medical applications. Behavior of the method is analyzed across variety of GPU and MIC (Intel Xeon Phi) accelerated clusters and results are presented.
Thesis Canvas: pdf