超声成像包括多种成像模式，B模式成像可以获得组织的结构信息，剪切波成像、矢量血流成像、超分辨率成像等新型成像模式可以获得组织的不同生理参数。上述新型成像模式依赖于较高的成像帧频，而传统聚焦波成像无法满足该要求。因此，高帧频成像已成为当前超声领域的研究热点。通常来说，成像帧频和图像质量是超声成像两个相互制约的参数。本文分别从超声成像的发射和接收波束合成两个角度，提出了针对高帧频成像的高质量图像重建方法，并通过算法设计，将图像重建时间控制在近实时水平。从超声成像发射的角度，递进式地提出了同相/正交数据域内的压缩感知-合成孔径成像（IQ CS-STA）、基于极小范数最小二乘求解的合成孔径成像（LS-STA）和基于深度学习数据恢复的合成孔径成像（CNN-STA），从而解决传统基于压缩感知的合成孔径成像（CS-STA）恢复STA数据集耗时（从数十分钟减少到数百毫秒）以及误差大（误差减小至三分之一）的问题。实验证明，经过以上三步研究，本文可以从较少次数平面波高帧频发射中准确并近实时地恢复出完整合成孔径数据，从而获得高质量的图像。从超声成像接收的角度，将超声波束合成建模为线性逆问题的求解过程，从而分离来自不同空间位置的回波信号，进而重建出更高质量的图像。引入深度学习技术进行该逆问题的快速求解时，常见的监督学习策略会面临无法获得活体组织真实散射系数（TRF）作为网络训练标签的问题。本文提出一种基于自监督学习的平面波图像重建方法，将系统采集的射频（RF）通道数据同时作为网络训练的输入和标签，绕开了训练过程对于TRF真值的需求。计算机仿真、仿体和在体实验证明，该方法大幅提高了图像的空间分辨率和对比度，并且相比于传统逆问题求解方法获得了3到4个数量级的重建加速。本文还进一步将该方法扩展成为了一个逆问题求解的一般框架，并在多个超声成像问题中取得了优于传统方法的效果。本文从发射和接收两个角度，提出了针对高帧频超声成像的高质量图像重建方法，并将重建时间大幅降低至近实时水平，从而为高帧频超声成像的临床应用提供了重要的研究基础。

Ultrasound imaging has become a widely used diagnostic tool to measure the various physiological parameters of the inspected tissue with B-mode imaging and many other new imaging modes, including shear wave elastography, vector flow imaging and super-resolution imaging. The conventional focused imaging could no longer meet the requirements of these new imaging modes for frame rate. Therefore, the development of a high-frame-rate imaging technique has become the current research hotspot of the ultrasound imaging field. Typically, there is a trade-off between the imaging frame rate and image quality. This study focuses on achieving high-frame-rate and high-quality ultrasound imaging at the same time by developing advanced and near-real-time image reconstruction algorithms.From the aspect of benefiting from the comprehensive advantages of different transmit sequences, we propose three methods to overcome the time-consuming (reduced from tens of minutes to hundreds of milliseconds) and large-recovery-error (reduced by two thirds) problems of the conventional compressed sensing based synthetic transmit aperture imaging (CS-STA) step by step. More specifically, we first extend the conventional CS-STA method from the radio-frequency (RF) domain to the in-phase/quadrature (IQ) domain to reduce the size of the dataset to recover. Thereafter, we propose a minimal l2-norm least squares method (LS-STA) to obtain the analytical solution to the STA dataset recovery problem to reduce the recovery time for each sample. Finally, the deep learning technique is implemented to reduce the recovery errors introduced by LS/CS-STA methods. Simulation, phantom and in vivo experiments demonstrate that this study can achieve high-frame-rate and high-quality STA imaging with high reconstruction speed.From the aspect of developing advanced beamforming algorithm, we propose to model the ultrasound beamforming as a linear inverse problem, to separate the echoes reflected from the different locations in RF data to reconstruct the tissue reflectivity function (TRF) for a better image quality. If we want to introduce deep learning techniques to achieve high-quality and high-speed reconstruction, the conventional supervised learning strategy would suffer from the lack of ground truth TRF (as training label) of the imaging target (which is impossible to obtain for in vivo tissue). Therefore, in this study, we propose a self-supervised learning strategy to utilize the acquired radio-frequency (RF) data as both the inputs and labels to the deep network, to bypass the requirements of ground truth TRF during the training process. Simulation, phantom and in vivo experiments demonstrate that the proposed method significantly improves the spatial resolution and contrast of the reconstructed images, and is three to four orders of magnitude faster than the state-of-the-art SR methods. This study further extends the proposed self-supervised learning strategy to a general framework to solve the other inverse problems in ultrasound imaging.This thesis improves the ultrasound image quality by combining different ultrasound sequences and developing advanced beamforming algorithm. In addition, the proposed methods have very high computational efficiencies and both can obtain ultrasound images in near-real-time. In conclusion, this thesis achieves much higher image quality than the conventional high-frame-rate imaging methods at the cost of a relatively small computational complexity, which paves the way for its real application in clinics.