A Conditional Adversarial Network for Single Plane Wave Beamforming

Compared to traditional focused transmissions, plane wave (PW) ultrasound imaging could enable higher frame rates by over a hundred fold, which is clinically relevant to real-time applications and ultrafast imaging. However, along with reducing acquisition time, PW imaging is confounded by image quality degradation, acoustic clutter, and speckle noise. To tackle this problem, we present a deep learning-based method to analyze raw radiofrequency (RF) channel data acquired by the ultrasound probe and convert this signal to the final B-mode image, bypassing the traditional beamforming procedure. The deep learning architecture relies on a conditional generative adversarial network (cGAN), in which the generative model and classifying model work simultaneously to produce an indistinguishable output from a ground truth. The cGAN was trained to predict B-mode images that look like beamformed PW results after multiple insonifications. This network was trained and tested utilizing a publicly accessible PICMUS database composed of in vivo and ex vivo ultrasound inclusions with randomly distributed scatterers in various combinations. The proposed method produces signal-to-noise ratio (SNR) enhancements from 1.112 to 1.540 when compared with conventional delay-and-sum(DAS) beamforming of a single PW insonification. The cross-correlation coefficient between a 75 plane wave image and cGAN-predicted data is 0.976, compared to 0.641 with DAS beamforming of a single PW insonification. These results demonstrate the feasibility of using this adversarial network to substitute traditional DAS beamforming in future applications.相较于传统聚焦式超声发射，平面波（plane wave, PW）成像可实现百倍级的帧速率提升，这对于临床实时应用与超快速成像场景具有重要意义。然而，伴随采集时间的缩短，平面波成像会面临图像质量退化、声学杂波与斑点噪声干扰的问题。为解决该难题，我们提出一种基于深度学习的方法，可分析超声探头采集到的原始射频（radiofrequency, RF）信道数据，并绕过传统波束成形流程，直接将该信号转换为最终的B型超声图像。该深度学习架构依托条件生成对抗网络（conditional generative adversarial network, cGAN），其中生成模型与分类模型协同工作，以生成与基准真值难以区分的输出。该cGAN被训练用于预测经多次平面波发射后的波束成形B型图像。我们利用公开可获取的PICMUS数据库完成网络的训练与测试，该数据库包含体内与体外超声仿体，其内部随机分布着不同组合的散射体。相较于仅单次发射平面波的传统延迟叠加（delay-and-sum, DAS）波束成形方法，本方法可实现1.112至1.540倍的信噪比增益。75幅平面波成像结果与cGAN预测数据间的互相关系数达0.976，而单次平面波发射的DAS波束成形结果仅为0.641。上述结果证明，该对抗网络在未来应用中替代传统DAS波束成形技术具备可行性。