Beamforming with deep learning from single plane wave RF data

Abstract

Deep learning approaches for improving ultrasound image reconstruction have proven successful in both experimental and clinical settings. In this paper, we present an autoencoder-based deep learning framework for ultrasound beamforming from the radio-frequency (RF) data received after a single plane wave transmission. Motivated by U-Net, the network consists of an encoder and a decoder. The network was trained and evaluated on simulated, phantom, and in vivo datasets. When tested on simulated data, the mean SNR, contrast, and gCNR of the learned image results were 3.16, -35.96 dB and 1.0 respectively, as well as a mean PSNR of 18.61 dB when compared to enhanced B-mode images. Each of these metrics outperformed the standard delay-and-sum (DAS) beamforming algorithm for the single plane wave image. In addition, the network was evaluated on an in vivo breast mass, achieving improved image quality compared to the corresponding single plane wave image. These results highlight the promise of exploring the proposed network to generate high quality ultrasound images from one plane wave, which could be applied to multiple ultrasound-based clinical tasks.