Use of B-splines in fast dynamic ultrasound RF simulations

Synthetic ultrasound data is invaluable for applications such as tuning image segmentation algorithms. Unfortunately, simulation time can become prohibitive for phantoms consisting of a large number of point-scatterers. In such cases, simplified simulation algorithms can be useful, and trade accuracy for speed. Storage of dynamic scatterer phantoms requires vast amounts of disk space when the number of time steps is large. We propose (1) a spline-based approach for efficient handling of dynamic scatterer phantoms with arbitrary time resolution, and (2) a fast, open source C++/CUDA implementation of a previously published convolution-based technique (COLE by Gao et al.) with support for our proposed spline-based approach. We observed a speedup of several orders of magnitude when comparing the simulation time for our CUDA implementation to the simulation time from an experiment reported in the literature, when using a similar simulation setup. We also observed that the spline-based representation of scatterers can improve the simulation speed almost four times for a dynamic simulation using the GPU, compared to the alternative of uploading new scatterer data to the GPU memory at every time step.