[论文解读] Deep Learning for Ultrasound Beamforming

Ruud J. G. van Sloun, Jong Chul Ye+1 more|arXiv (Cornell University)|Sep 23, 2021

Ultrasound Imaging and Elastography被引用 4

一句话总结

本文提出一种基于自适应实例归一化（AdaIN）层的可切换深度波束成形器，将多种超声波束成形方法（如DAS、MVBF和反卷积）统一于单一神经网络中。通过使用可学习的AdaIN码对网络进行条件控制，可在不重新训练的情况下生成多种波束成形输出，显著降低内存开销，同时在图像质量上优于传统方法。

ABSTRACT

Diagnostic imaging plays a critical role in healthcare, serving as a fundamental asset for timely diagnosis, disease staging and management as well as for treatment choice, planning, guidance, and follow-up. Among the diagnostic imaging options, ultrasound imaging is uniquely positioned, being a highly cost-effective modality that offers the clinician an unmatched and invaluable level of interaction, enabled by its real-time nature. Ultrasound probes are becoming increasingly compact and portable, with the market demand for low-cost pocket-sized and (in-body) miniaturized devices expanding. At the same time, there is a strong trend towards 3D imaging and the use of high-frame-rate imaging schemes; both accompanied by dramatically increasing data rates that pose a heavy burden on the probe-system communication and subsequent image reconstruction algorithms. With the demand for high-quality image reconstruction and signal extraction from less (e.g unfocused or parallel) transmissions that facilitate fast imaging, and a push towards compact probes, modern ultrasound imaging leans heavily on innovations in powerful digital receive channel processing. Beamforming, the process of mapping received ultrasound echoes to the spatial image domain, naturally lies at the heart of the ultrasound image formation chain. In this chapter on Deep Learning for Ultrasound Beamforming, we discuss why and when deep learning methods can play a compelling role in the digital beamforming pipeline, and then show how these data-driven systems can be leveraged for improved ultrasound image reconstruction.

研究动机与目标

• 解决在临床超声系统中为每种波束成形方法（如DAS、MVBF）单独训练深度学习模型所带来的高内存和高计算成本问题。
• 实现单一深度学习模型通过条件AdaIN层生成多种波束成形输出（如DAS、MVBF、去斑图像）。
• 通过探索CycleGAN等无监督学习策略，减少对成对训练数据的依赖，以实现图像质量增强。
• 在保持实时性能的同时，利用数据驱动的波束成形方法提升超声图像的保真度、分辨率和对比度。

提出的方法

• 使用成对输入-目标数据，通过监督学习训练深度神经网络，将原始超声通道数据映射为波束成形图像。
• 网络引入自适应实例归一化（AdaIN）层，通过从目标波束成形方法中提取的风格码对特征图进行条件控制。
• AdaIN变换通过重加权特征图的均值和方差，使其匹配参考风格（如DAS或MVBF）的统计特性，从而实现输出的动态切换。
• AdaIN码由轻量级网络生成，可在不存储多个完整模型的情况下实现高效推理。
• 对于无监督学习，采用基于CycleGAN的无配对图像翻译方法，利用来自不同来源的高质量图像，从低端系统中提升图像质量。
• 通过仅更改AdaIN码，系统可实现波束成形的可切换性，支持在不同成像设置下的实时自适应。

Figure 1: An illustration of three transmit types. Transmit events are denoted by $e$ , receive channels as $c$ and, for focused scanlines, $2L+1$ is the size of the active aperture in terms of elements.

实验结果

研究问题

• RQ1 是否可以训练单一深度神经网络，使其在不为每种方法单独训练模型的情况下，有效模拟多种波束成形技术（如DAS、MVBF）？
• RQ2 AdaIN层在条件控制深度波束成形器以生成多样化波束成形输出方面，效果如何？
• RQ3 当缺乏真实配对数据时，无监督学习对波束成形质量的影响如何？
• RQ4 统一的波束成形器是否能在保持或提升图像质量的同时，显著降低临床超声系统中的内存和计算成本？

主要发现

• 采用AdaIN层的单一深度波束成形器，通过不同AdaIN码的条件控制，成功生成多种波束成形输出（如DAS、MVBF、反卷积、去斑图像）。
• 与为每种波束成形方法单独训练模型相比，基于AdaIN的方法显著降低了内存开销，有利于在便携式超声系统中实际部署。
• 通过CycleGAN实现的无监督学习可利用无配对数据提升图像质量，证明了在真实场景中缺乏真实配对数据时的可行性。
• 该方法在传统波束成形技术基础上，显著提升了图像对比度和分辨率，尤其在低信噪比（low-SNR）和低帧率条件下表现更优。

Figure 2: General beamforming scheme.