Beamforming with deep learning from single plane wave RF data

Deep learning approaches for improving ultrasound (US) 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 of a single plane wave. Motivated by U-Net, the network consists of an encoder and a decoder. The network is trained and evaluated with simulated and \textit{in vivo} datasets. When tested on simulated data, the beamformed images from the proposed network achieved an SNR of 25.02, contrast of -40.61dB, and gCNR of 1.0, a mean PSNR of 18.64dB compared to ground truth images and mean axial and lateral FWHM of 0.42 and 0.61 respectively. Each of these metrics outperformed the standard delay and sum (DAS) beamforming algorithm. In addition, the network was evaluated on an \textit{in vivo} breast mass, achieving improved SNR of 3.81. These results demonstrate that the proposed network is capable of generating high quality US images from only one plane wave, which could be applied to multiple ultrasound-based clinical tasks.用于优化超声（US）图像重建的深度学习方法已在实验与临床场景中被验证具有优异效果。本文提出一种基于自编码器的深度学习框架，可基于单平面波的射频（RF）数据完成超声波束形成。该网络借鉴U-Net架构，由编码器与解码器组成。研究团队利用仿真数据集与体内（in vivo）数据集对该网络进行训练与评估。在仿真数据测试环节，相较于真值图像，所提网络生成的波束形成图像信噪比（SNR）为25.02、对比度为-40.61dB、广义对比度噪声比（gCNR）为1.0，平均峰值信噪比（PSNR）达18.64dB，轴向与横向半高全宽（FWHM）分别为0.42与0.61。上述所有指标均优于标准延迟叠加（DAS）波束形成算法。此外，该网络在体内乳腺肿块数据集上完成评估，实现了3.81的信噪比提升。上述结果表明，所提网络仅需单平面波即可生成高质量超声图像，可应用于多种基于超声的临床任务。