Mechanical characterization of tissue-like materials using information based machine learning

Ultrasound;Neural Networks;Elasticity;Neural Network Constitutive Model;Finite-element Analysis

Hoerig, Cameron Lee ; Insana ; Michael F.

关键词: Ultrasound;    Neural Networks;    Elasticity;    Neural Network Constitutive Model;    Finite-element Analysis;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/89003/HOERIG-THESIS-2015.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Changes in the mechanical properties of soft tissues may be indicative of disease processes. Medical elastography techniques are an attempt to create images of the mechanical behavior to increase the sensitivity and specificity of existing imaging modalities. Current quantitative elasticity imaging methods rely on a priori assumptions of the tissue biomechanics in order to simply the forward problem, from which an inverse problem is developed. Erroneous assumptions and noisy image data result in incorrect estimates of the mechanical parameters.This thesis presents a new method of characterizing the mechanical response of soft tissues. Machine-learning techniques and measured force-displacement data are used to create empirical models of the constitutive behavior. Informational models are developed without enforcing simplyfing assumptions of the true underlying mechanics, allowing the mechanical properties of the tissue to be investigated after the model is developed. Knowledge of the true behavior allows the appropriate consitutive model to be chosen to create a parametric summary of the tissue suitable for imaging.The informational modeling process is demonstrated on gelatin phantoms comprised of a soft background material with one or three stiffer inclusions. An ultrasound probe was used to uniaxially compress the phantoms while acquiring surface force and displacement data, as well as ultrasound images. A speckle-tracking algorithm estimated motion of the phantoms within the imaged region. Force-displacement data and the Autoprogressive training algorithm was then used to build informational models describing the constitutive behavior of the gelatin materials. It will be shown that estimates of the full stress and strain vectors throughout an entire model can be computed with the use of informational models, a feat not previously possible in ultrasound elastography. These vectors can then be used to create a parametric summary of mechanical properties of the gelatin materials - in this case, estimates of the Young's modulus. The resuling images of the Young's modulus distribution clearly differentiate the stiff inclusion(s) from the soft background.Results from this investigation are just the starting point for developing informational models of soft tissues. Sampling requirements and training methods to improve the ability of the models to characterize the linear-elastic properties of the gelatin are discussed. Future work will involve extending this method to 3D and characterizing more complex mechanical behaviors, including nonlinear, time-dependent, path-dependent properties.