Deep Learning-Based Ultrasound Beam Shaping for Spatiotemporal Acoustic Holograms Generation

Abstract

Acoustical hologram generation can be achieved via controlled beam shaping by engineering the transmitted phases to create a desired pattern. Optically inspired phase retrieval algorithms and standard beam shaping methods assume continuous wave (CW) insonation, which successfully generate acoustic holograms for therapeutic applications that involve long burst transmissions. However, a phase engineering technique designed for single-cycle transmission and capable of achieving spatiotemporal interference of the transmitted pulses is needed for imaging applications. Toward this goal, we developed a multilevel residual deep convolutional network for calculating the inverse process that will yield the phase map for the creation of a multifoci pattern. The ultrasound deep learning (USDL) method was trained on simulated training pairs of multifoci patterns in the focal plane and their corresponding phase maps in the transducer plane, where propagation between the planes was performed via singe cycle transmission. The USDL method outperformed the standard Gerchberg-Saxton (GS) method, when transmitted with single cycle excitation, in parameters including the number of focal spots that were generated successfully and their pressure and uniformity. In addition, the USDL method was shown to be flexible in generating patterns with large focal spacing, uneven spacing, and nonuniform amplitudes. In simulations, the largest improvement was obtained for four foci patterns, where the GS method succeeded in creating 25% of the requested patterns, while the USDL method successfully created 60% of the patterns. These results were confirmed experimentally via hydrophone measurements. Our findings suggest that deep learning-based beam shaping can facilitate the next generation of acoustical holograms for ultrasound imaging applications.