Machine learning, finite element analysis and quasi-static ultrasound nonlinear elastography

Wang, Yiliang

Permalink

https://hdl.handle.net/2142/117752 Copy

Description

Title

Machine learning, finite element analysis and quasi-static ultrasound nonlinear elastography

Author(s)

Wang, Yiliang

Issue Date

2022-11-14

Director of Research (if dissertation) or Advisor (if thesis)

Insana, Michael F.

Doctoral Committee Chair(s)

Hilgenfeldt, Sascha

Committee Member(s)

• Saif, Taher A.
• Ghaboussi, Jamshid

Department of Study

Mechanical Sci & Engineering

Discipline

Mechanical Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Machine learning
• Finite element analysis
• Ultrasound imaging
• Elastography.

Abstract

Quasi-static ultrasound elastography (QUSE) is a technique which has been used to diagnose and study pathological soft tissues. The Autoprogressive method (AutoP) which is an inverse algorithm that combines machine learning (ML) and finite element analysis (FEA) has been applied in the reconstruction of QUSE. Previous applications of AutoP in QUSE mainly focused on linear elastic mechanical behaviors of the target media because the target is under infinitesimal deformation. The objective of this dissertation is to leverage AutoP to reconstruct nonlinear elastography of soft materials that undergo large deformations from ultrasonic imaging data. Since FEAs are iteratively executed in AutoP and its speed affects the entire computational cost, we developed a fast and user-friendly self-designed FEA solver specifically for the AutoP. To achieve this goal, on the one hand, the solutions of FEA in the previous AutoP iteration are utilized to initialize current FEA to reduce the number of Newton-Raphson iterations. On the other hand, we managed to use this self-designed FEA solver to execute multiple different FEA tasks in parallel. Meanwhile, in order to avoid the numerical instability induced by the incompressbility constraints and tackle frictionless contact problems, a mixed finite element formulation and a segment-to-segment algorithm are used in this self-designed FEA solver, respectively. A number of benchmark tests are conducted using the self-designed FEA solver and the commercial FEA solver ABAQUS and the calculated results from these two solvers are in a good agreement. In addition to the aforementioned advantages of the self-designed FEA solver, a C++ class for the neural network constitutive model (NNCM) is designed in this FEA solver, which can benefit the further development of new architectures of NNCM. With the self-designed FEA solver, we extended AutoP to characterize nonlinear elastic mechanical properties of a heterogeneous object with known internal geometries. Previous versions of NNCM utilize strain-stress form of input and output and hence have high training data dimensionality which might induce difficulties in AutoP training. By leveraging unique properties of nonlinear elastic media, we first created a new architecture of NNCM to facilitate the AutoP learning. In the meantime, the framework of AutoP was adapted to be able to train this new NNCM. Further, we created and investigated a new stopping rule for the NNCM to prevent it from being overtrained. Finally, the AutoP training integrated with the new NNCM was validated using both synthetic and experimental data recorded from 3D objects and it was found that the results were robust to measure errors and spatial variations in mechanical properties. In the last part of this dissertation, we converted the AutoP with this NNCM into a true imaging tool which can be used to visualize the 3D nonlinear elastography. First, based on the current NNCM, we created a nonlinear Cartesian neural network (NonlCaNNCM) which can describe both the mechanical properties and the corresponding spatial distribution. Then, we developed a training scheme which can help AutoP to reconstruct 3D nonlinear elastography by using multiple 2D measurement datasets. This methodology is then validated using both synthetic and experimental data. It also discusses about the factors which can affect the performance of the training. It is our hope that the techniques developed in this dissertation will eventually provide means for clinicians into making better diagnosis and designing better treatment plans for breast cancer.

Graduation Semester

2022-12

Type of Resource

Thesis

Copyright and License Information

Copyright 2022 Yiliang Wang

Owning Collections