High Performance Computing in Ultrasound Cancer Treatment

Ultrasound simulation, high performance computing, spectral methods, medical applications.

According to the Czech Society of Oncology, more than 73,000 tumour diseases are newly diagnosed in the Czech Republic every year and this number is continuing to grow. A very promising alternative to the standard treatment procedures is a non-invasive high intensity focused ultrasound (HIFU), also known as focused ultrasound surgery. The critical component of the effective HIFU treatment is the preoperative treatment planning to precisely place the focus at a desired position and determine an appropriate dosage. However, modelling the ultrasound propagation in human body is both physically complex and computationally challenging. This thesis summarises original research in the area of large-scale acoustic model development,  I have been involved in since 2011. During this period, my colleagues and I have published a large number of scientific papers, fourteen of which are presented in this thesis. First, the acoustic model accounting for combined effect of heterogeneity, non-linearity and absorption is discussed. Next, the numerical solution using a corrected k-space pseudo-spectral method is introduced, and its main benefits and drawbacks are discussed. Consequently, the implementation, validation and performance of large-scale distributed simulation codes are outlined. The focus is put on two approaches of the simulation domain decomposition, namely the global domain decomposition with a superior accuracy but inherited communication bottleneck, and the novel local Fourier basis decomposition making a compromise between the numerical accuracy and the communication overhead. The performance, scaling and simulation cost of the developed simulation codes are evaluated on best supercomputers with thousands of processor cores and hundreds of graphics processing units. Finally, several clinical applications in the prostate, kidney and brain are presented and their impact discussed. The accurate acoustic model and distributed simulation codes presented in this work have opened the door for disruptive science by allowing to simulate HIFU in domains comprising more than 8,000cm^3 in a clinically meaningful time. This constitutes more than 250x bigger simulation domain than ones commonly used at the time I joined this project.