Switchable deep beamformer for high-quality and real-time passive acoustic mapping

TL;DR

This work addresses the need for real-time, high-quality passive acoustic mapping (PAM) of cavitation during ultrasound therapy by introducing a switchable deep beamformer built on a generative adversarial network. The network maps raw RF signals directly to PAM images and can operate across linear and phased arrays (1–15 MHz) by incorporating an array mask as prior information, achieving image quality comparable to EISRCB but with drastically reduced computational cost. Training uses a hybrid dataset of simulated and experimental data, with losses designed to optimize realism, reconstruction accuracy, and cross-array consistency. The results show substantial improvements in energy-spread reduction and artifact suppression, real-time reconstruction speeds on GPUs, and robust localization across noise levels and experimental conditions, indicating strong potential for real-time cavitation monitoring in ultrasound therapy.

Abstract

Passive acoustic mapping (PAM) is a promising tool for monitoring acoustic cavitation activities in the applications of ultrasound therapy. Data-adaptive beamformers for PAM have better image quality compared to the time exposure acoustics (TEA) algorithms. However, the computational cost of data-adaptive beamformers is considerably expensive. In this work, we develop a deep beamformer based on a generative adversarial network, which can switch between different transducer arrays and reconstruct high-quality PAM images directly from radio frequency ultrasound signals with low computational cost. The deep beamformer was trained on the dataset consisting of simulated and experimental cavitation signals of single and multiple microbubble clouds measured by different (linear and phased) arrays covering 1-15 MHz. We compared the performance of the deep beamformer to TEA and three different data-adaptive beamformers using the simulated and experimental test dataset. Compared with TEA, the deep beamformer reduced the energy spread area by 18.9%-65.0% and improved the image signal-to-noise ratio by 9.3-22.9 dB in average for the different arrays in our data. Compared to the data-adaptive beamformers, the deep beamformer reduced the computational cost by three orders of magnitude achieving 10.5 ms image reconstruction speed in our data, while the image quality was as good as that of the data-adaptive beamformers. These results demonstrated the potential of the deep beamformer for high-resolution monitoring of microbubble cavitation activities for ultrasound therapy.

Figures (7)

• Figure 1: Distribution of (a) one bubble cloud and (b) two bubble clouds in the simulations. (c) Schematic diagram of the experimental setup. (d) The phantom with an inclined tube embedded. Experimental setup 1 (e) and 2 (f) of multiple bubble clouds for tubes. (g) Experimental setup of FUS treatment for mice and (h) the sequence for FUS treatment and data acquisition for B-mode imaging and PAM.
• Figure 2: Architecture of the switchable deep beamformer for PAM: (a) the generator, (b) the discriminator/classifier. The dimension below each blocks denotes the size of the output feature maps. k, s, p denotes the number of kernel size, stride and size of padding, respectively. LAA: local aware attention; PA: pixel attention.
• Figure 3: (a) Simulated and (b) in vitro experimental PAM images reconstructed by the EISRCB, RLPB, DAX-RCB and TEA for P4-1. Cumulative probability distribution of the Euclidean distance between the center of the bubble clouds and source position localized by (c) EISRCB, (d) RLPB, (e) DAX-RCB and (f) TEA.
• Figure 4: PAM images reconstructed by the deep beamformer, EISRCB, and TEA using the simulated test set of (a--c) single and (d--f) two bubble clouds for P4-1, L7-4 and CL15-7. Cumulative probability distribution of lateral position deviation and axial position deviation between the deep beamformer yielded images and EISRCB images for (g--h) single source and (i--j) multiple sources. The wavelength ($\lambda$) was evaluated at the center frequency of each transducer array.
• Figure 5: PAM images reconstructed by the deep beamformer and EISRCB using the simulated test set for L7-4 with noise of 15 dB noise and 0 dB.