Abstract
Traditional ultrasound field simulations rely on the Spatial Impulse Response (SIR) method, which often suffers from geometric approximation errors and high computational costs due to extreme sampling requirements. This paper presents a novel simulation framework that utilizes Non-Uniform Rational B-Splines (NURBS) for exact transducer geometry representation and high-order B-spline bases for the temporal signal. By leveraging the bandlimited nature of ultrasound pulses, the proposed method achieves high accuracy at sampling rates near the Nyquist limit, offering potential speedups of up to two orders of magnitude over conventional CPU-based simulators.
Introduction
In ultrasound imaging, calculating the pressure field generated by a transducer is fundamental. The gold-standard approach is the Spatial Impulse Response (SIR) method, which computes a time-dependent surface integral over the transducer face. However, two major bottlenecks persist:
- Geometric Discretization: Complex transducer shapes are typically approximated with thousands of planar elements (rectangles or triangles), leading to geometric errors and high memory usage.
- Sampling Requirements: The SIR contains high-frequency discontinuities at the arrival times of waves from the transducer edges. To avoid aliasing, simulators must often sample at 10–100 times the center frequency of the ultrasound pulse, significantly increasing computational load.
Methodology
The proposed pipeline addresses these issues through a unified spline-based approach that optimizes both geometric representation and temporal sampling.
Exact Geometry with NURBS
Instead of mesh-based approximations, the simulator represents transducer surfaces using Non-Uniform Rational B-Splines (NURBS). NURBS provide a mathematically exact representation of common shapes like spheres and cylinders. These surfaces are decomposed into smooth Bézier patches, where high-order Gaussian quadrature rules are applied to evaluate the surface integral efficiently with very few points compared to uniform discretization.
B-spline Time Basis and Generalized Interpolation
The core innovation lies in representing the SIR as a sum of shifted-and-weighted B-spline basis functions. By treating the simulation in a continuous-time framework and using a "pre-filtering" step (via IIR filters) on the excitation pulse, the final ultrasound signal is obtained through a discrete convolution between pulse coefficients and a "basis SIR." This framework effectively exploits the bandlimited nature of ultrasound pulses.
Results
The method was evaluated using synthetic Dirac streams and analytic benchmarks for spherically focused and rectangular elements.
Accuracy and Convergence
The experiments demonstrated that the relative error decreases according to the "degree + 1" rule for B-splines. For instance, a 5th-degree B-spline achieved a relative error of $10^{-4}$ at a low 30 MHz sampling rate and dropped to $10^{-7}$ at 80 MHz. This high accuracy at low sampling rates is a significant improvement over standard linear or nearest-neighbor interpolation methods.
Computational Efficiency
By reducing the required sampling rate to near the Nyquist limit of the excitation pulse, the computational burden is drastically lowered. Initial tests suggest that a GPU-accelerated implementation of this spline-based method could be two orders of magnitude faster than industry-standard CPU simulators like Field II, while simultaneously eliminating common low-sampling artifacts such as "ripple" distortion.
Conclusion
The integration of NURBS for geometry and high-order B-splines for temporal representation provides a robust, accurate, and efficient framework for ultrasound simulation. Future work will focus on extending this model to inhomogeneous media and further optimizing quadrature point selection.
Summary updated