Bioinformatics and Computational Biology, University of Minnesota, Rochester, MN, USA; Department of Radiology, Mayo Clinic, Rochester, MN, USA.
AGH University of Science and Technology, Krakow, Poland.
Comput Biol Med. 2021 Jun;133:104382. doi: 10.1016/j.compbiomed.2021.104382. Epub 2021 Apr 11.
Ultrasound shear wave elastography (SWE) techniques have been very useful for the analysis of tissue rheological properties, but there are still obstacles for robust evaluation of viscoelastic tissue properties. In this proof-of-concept study, we investigate whether convolutional neural networks (CNN) are capable of retrieving the elasticity and viscosity parameters from simulated shear wave motion images. Staggered-grid finite difference simulations based on a Kelvin-Voigt rheological model were used to generate data for this study. The wave motion datasets were created using Kelvin-Voigt shear elasticity values ranging from 1 to 25 kPa, shear viscosities ranging from 0 to 10 Pa⋅s, and two different push profiles using f-numbers of 1 and 2. The CNN architectures, optimized using mean squared error loss, were then trained to retrieve a specific viscoelastic parameter. Both elasticity and viscosity values were successfully retrieved, with regression R values above 0.99 when correlating the estimated mechanical properties versus the true mechanical properties. The CNN performance was also compared to estimation of shear elasticity and viscosity from fitting dispersion curves estimated from two-dimensional Fourier transform analysis. The results demonstrated that the CNN models were robust to noise, vertical position and partially to f-number. The architecture was proven to be robust to multiple push profiles if trained properly. The CNN results showed higher accuracy over the full viscoelastic parameter range compared to the Fourier-based analysis. The overall results showed the CNNs' potential to be an alternative to complex mathematical analyses such as Fourier analysis and dispersion curve estimation used currently for shear wave viscoelastic parameter estimation.
超声剪切波弹性成像（SWE）技术在分析组织流变性质方面非常有用，但对于稳健评估粘弹性组织性质仍然存在障碍。在这项概念验证研究中，我们研究了卷积神经网络（CNN）是否能够从模拟的剪切波运动图像中检索弹性和粘性参数。基于 Kelvin-Voigt 流变模型的交错网格有限差分模拟用于生成本研究的数据。使用 Kelvin-Voigt 剪切弹性值范围为 1 至 25kPa、剪切粘度范围为 0 至 10Pa⋅s 以及使用 f-数为 1 和 2 的两种不同的推压轮廓，创建了波运动数据集。使用均方误差损失优化的 CNN 架构，然后被训练以检索特定的粘弹性参数。当将估计的力学性质与真实的力学性质相关联时，弹性和粘度值都成功地被检索到，回归 R 值大于 0.99。还将 CNN 性能与从二维傅里叶变换分析估计的频散曲线拟合估计的剪切弹性和粘度进行了比较。结果表明，CNN 模型对噪声、垂直位置以及部分对 f-数具有鲁棒性。如果正确训练，该架构被证明对多种推压轮廓具有鲁棒性。与基于傅里叶的分析相比，CNN 结果在全粘弹性参数范围内显示出更高的准确性。总体结果表明，CNN 有可能替代当前用于剪切波粘弹性参数估计的复杂数学分析，如傅里叶分析和频散曲线估计。
