How to inform a machine learning method of mechanical property imaging
Abstract: Our group has been developing a method of imaging the mechanical properties of soft tissues by combining ultrasonic measurements, finite-element physics and machine learning (the autoprogressive method). The process involves propagating measured force and displacement of an ultrasound probe and measured internal displacements into stress-strain training data via finite-element methods. The difficulty is that the measured data lacks necessary internal force information and internal displacements are only measured in two-dimensional scan planes while the experiment is three-dimensional. Further, improper modeling of boundary conditions of the experiment can introduce biases in the training data. Given the difficulties of this ill-posed, inverse problem, how can we guide networks to learn the correct mechanical behavior? I have been interested in answering this question and more generally understanding how networks learn given the information we provide them. In this talk, I will describe how machines learn through the lens of the autoprogressive method of mechanical property imaging.
Bio: I am a third year Ph.D student working in Dr. Michael Insana's lab. I received a bachelors degree in Physics from Rhodes College in Memphis, TN in 2021. My primary research interest is in developing data-driven methods for ultrasonic elasticity imaging,
