A data-driven approach to quasi-static ultrasonic elasticity imaging

Elasticity imaging;Inverse problems;Machine learning;Finite element analysis;Artificial neural networks

Hoerig, Cameron

关键词: Elasticity imaging;    Inverse problems;    Machine learning;    Finite element analysis;    Artificial neural networks;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/102440/HOERIG-DISSERTATION-2018.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 that accompany pathology can potentially be used as a biomarker for detection and diagnosis of disease. Quasi-static ultrasonic elastography (QUSE) is one of several elasticity imaging techniques sensitive to mechanical contrast and offers a way to visualize the spatiotemporal distribution of material properties within tissues. Unfortunately, QUSE is generally an ill-posed inverse problem. Quantifying the mechanical properties requires measurements of more stress-strain data than can be acquired during a typical ultrasonic (US) imaging exam. Model-based inverse methods attempt to circumvent these limitations in part by estimating the spatial distribution of a pre-defined set of material parameters. As a consequence, model-based methods provide no means for discovering new diagnostically-relevant mechanical properties or for exploring ranges of known model parameters for relevance in a given situation.We are developing a data-driven approach for quantitative QUSE using the Autoprogressive method (AutoP), which combines artificial neural networks (ANNs) and finite element analysis (FEA). AutoP has previously been used in geotechnical and civil engineering applications to build "soft-computational" models of materials. Using knowledge of object shape and force-displacement measurements, investigators were able to build neural network constitutive models (NNCMs) that accurately describe the behavior of linear, non-linear, and time-dependent materials with no prior constitutive model assumptions. Furthermore, NNCMs provide a means to estimate spatiotemporal stress and strain distributions from force-displacement data. NNCMs and AutoP offer a fundamentally different approach to QUSE. We first demonstrate that a very sparse sampling of force-displacement data is sufficient for estimating the linear-elastic properties of gelatin phantoms when the interior geometry is known. Then, we introduce Cartesian NNCMs (CaNNCMs), a novel ANN architecture, capable of learning both material property and geometric information. We begin exploring the spatial sampling requirements to reconstruct Young's modulus distributions in both 2-D and 3-D. Moreover, we show how CaNNCMs can be used to estimate the spatial distribution of all stresses and strains and can be directly interrogated to infer the mechanical properties governing measured data. Further development of this method to non-linear and viscoelastic materials may provide a means to discover the mechanical parameters most relevant to clinical elastography.