Labeled numerical phantom of abdominal wall for wave-physics based ultrasound imaging: applications to image reconstruction and parameter estimation (Part D)

The abdominal wall consists of an acoustically complex organization of tissue layers that generate significant degradation in diagnostic ultrasound imaging of internal organs. As with any non-invasive diagnostic medical imaging modality, the underlying ground truth is a fundamentally unknown quantity in a living patient. A realistic labeled 3D numerical phantom of the human abdominal wall is used to establish the acoustical properties of human tissue. Simulations based on the first principles of wave propagation, including the effects of propagation, aberration, reflection, refraction, attenuation, scattering, reverberation, and nonlinearity are used to generate a large data set of raw ultrasound data for 2D imaging. This data set is then used to train a physics-informed neural network to generate local sound speed estimates. Finally the translation of this approach is demonstrated in vivo on clinical data of liver ultrasound images. It is shown that together, the neural network and this training data set generate high quality estimates of the sound speed as measured by improvements in clinical image quality. The labeled abdominal data set, the simulation tools that model wave propagation, and the neural network approach are made publicly available. Other training and optimization approaches can be applied to this data.

人体腹壁由声学结构复杂的组织层构成，此类组织层会对内脏器官的诊断超声成像（diagnostic ultrasound imaging）造成显著的信号衰减。与所有非侵入式诊断医学成像模态（non-invasive diagnostic medical imaging modality）一样，活体患者体内的基础真值（ground truth）本质上是未知量。本研究采用经标注的真实感人体腹壁三维数值体模（3D numerical phantom），以此确定人体组织的声学特性（acoustical properties）。基于波传播（wave propagation）第一性原理开展仿真，涵盖传播、像差（aberration）、反射（reflection）、折射（refraction）、衰减（attenuation）、散射（scattering）、混响（reverberation）与非线性（nonlinearity）等效应，以此生成面向二维成像（2D imaging）的大规模原始超声数据集（raw ultrasound data）。将该数据集用于训练物理感知神经网络（physics-informed neural network），以实现局部声速估计（local sound speed estimates）。最终，本研究在肝脏超声图像的临床数据上开展体内（in vivo）实验，验证了该方法的实际效果。研究表明，将该神经网络与本训练数据集相结合，可通过提升临床图像质量（clinical image quality）实现高精度的局部声速估计。本研究将经标注的腹壁数据集、波传播建模仿真工具以及该神经网络方法公开共享，其他训练与优化方案均可基于本数据集开展研究。