BME Seminar: Ali Kamali
Monday, March 27, 2023 - 12:00 p.m.
Ali Kamali
Ph.D. Candidate
Biomedical Engineering
"Elasticity Imaging Using Physics-based Deep Learning: Spatial Discovery of Mechanical Properties"
Keating 103 | Live Zoom Link; Passcode: BearDown
(Instructor permission required for enrolled students to attend via Zoom)
Hosts: Dr. Beth Hutchinson and Dr. Shang Song
Persons with a disability may request a reasonable accommodation by contacting the Disability Resource Center at 621-3268 (V/TTY).
Abstract: Soft biological tissues often have complex mechanical properties due to spatial variation in structural components or fiber alignments. Due to these complexities, conventional mechanical characterization approaches rely on isolating small tissue specimens to minimize these spatial variations to more homogeneous estimates. As a less invasive characterization method, strain-based or quasi-static elastography is an imaging technique that reconstructs the spatial distribution of mechanical properties using deformation and force measurements. However, existing methods face challenges in accurately estimating the mechanical properties due to the ill-posed inverse mathematical problem.
We propose physics-based deep learning frameworks to estimate the spatial distribution of mechanical parameters in two and three dimensions using acquired deformation images and available mechanical stress boundary conditions. These models estimate distributions of the unknown variables (mechanical properties) and attempt to minimize loss functions comprised of governing physical equations that contain those variables. First, we developed parameter estimation physics-informed neural networks (PINNs) to solve inverse isotropic linear elasticity problems in two dimensions. We validated this model against data from experiments and numerical simulations to show its robustness over a range of complexities. These examples included experimental and simulated stiff inclusions in soft backgrounds, representing tumorous tissue, and a simulated human brain slice under loading that contained distinct gray and white matter regions. Next, we focused on 3D parameter estimation in isotropic and transversely isotropic linear elastic materials. To address this increased complexity, we updated our deep learning model to a UNet architecture. In addition, we proposed multiple innovative approaches, including spatial loss weighting, constrained loss terms, and training over multiple loading states, to ensure and speed up convergence to accurate estimations of mechanical parameter distributions. Our frameworks can be utilized in a wide range of engineering, biomedical, and clinical applications to discover mechanical properties inside a domain in non-invasive or minimally invasive ways.