Guiding the development of data-driven models to solve constitutive inverse problems in medical elasticity imaging

Newman, Will

Permalink

https://hdl.handle.net/2142/129833 Copy

Description

Title

Guiding the development of data-driven models to solve constitutive inverse problems in medical elasticity imaging

Author(s)

Newman, Will

Issue Date

2025-07-01

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

Insana, Michael

Doctoral Committee Chair(s)

Insana, Michael

Committee Member(s)

• Oelze, Michael
• Anastasio, Mark
• Lam, Fan

Department of Study

Bioengineering

Discipline

Bioengineering

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Elastography
• Constitutive Inverse Problems
• Ultrasound
• Machine Learning

Abstract

Elasticity imaging is an approach similar to manual palpation that replaces fingertip sensing with images that describe mechanical contrast between tissues. Elasticity images provide physicians with diagnostic information regarding the state of disease in many soft tissues. However, 3D quantitative images of tissue material properties are required to connect tissue-level images to cellular mechanobiology. Those seeking quantitative images of material properties formulate medical elasticity imaging as a 3D constitutive inverse problem. Ideally, measurements of force and displacement would propagate into a complete set of stresses and strains at all points in the medium to invert the constitutive equation and estimate material properties in a volume. However, force can only be measured on the tissue surface, and displacement is typically measured in a 2D plane. How can we propagate sparse measurements of surface force and planar displacement into a set of stresses and strains that can uncover quantitative material properties of tissues? This thesis has spearheaded the development of an ultrasonic-based technique for elasticity imaging throughout a tissue volume using the autoprogressive (AutoP) method. AutoP combines object-specific measurements with finite-element analysis (FEA) to propagate sparse measurements into the stresses and strains needed to uncover constitutive behavior in the tissue volume. A machine learning framework learns spatially varying constitutive behavior from the stresses and strains and replaces the constitutive matrix within FEA. The goal is to provide AutoP with measurements that lead to a physically-consistent, data-driven solution that adequately captures the deformation of the material. In this dissertation, I examine the features that influence model development in AutoP. The focus is on the creation of efficient training strategies and measurement acquisition techniques that can generate a rich training environment. I developed new metrics that can monitor how well the measurements and training parameters contribute to the learning process. Factors that degrade the image quality, specifically those associated with experimental acquisition of measurements, were quantified through measures of spatial resolution and contrast. Knowledge of the image quality metrics led to the validation of this technique using a variety of simulated and manufactured gelatin phantoms targeting BI-RADS features of suspicious breast lesions. This work develops the image science of the AutoP method through rigorous examination of the learning process. The product of this work can be described as a users manual for proper use of AutoP as a tool for imaging the quantitative material properties of tissues in 3D. The scientific developments in this dissertation were a necessary step toward establishing AutoP as a robust tool for diagnostic imaging and mechanobiological discovery.

Graduation Semester

2025-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129833

Copyright and License Information

Copyright 2025 Will Newman

Owning Collections