Modeling a Clinical Acoustic Information System Using Physics-Informed Machine Learning
Open Access
- Author:
- Alkhadhr, Shaikhah
- Graduate Program:
- Computer Science and Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- September 29, 2023
- Committee Members:
- Chitaranjan Das, Program Head/Chair
Mohamed Almekkawy, Chair & Dissertation Advisor
Kamesh Madduri, Major Field Member
David Koslicki, Major Field Member
Yun Jing, Outside Unit & Field Member - Keywords:
- physics-informed neural networks
wave equation
ultrasound therapeutics
numerical modeling
information systems - Abstract:
- Numerical modeling is one of the most important methods used for understanding and developing physical systems in different scientific and industrial fields. Particularly in the simulation of ultrasound waves, modeling linear and nonlinear wave equations can endure several complications, including the Curse of Dimensionality (CoD), exhaustion of computational resources, and the development of discontinuities in finite time. In the machine-learning sector, Deep Neural Networks (DNNs) have gained a noticeable acceleration of development in the last decade, branching into applications in almost every professional and academic field. This introduced a new architecture for DNNs to solve Partial Differential Equations (PDEs). Physics-Informed Neural Networks (PINNs) are capable of predicting the PDE solution that governs a physical phenomenon. The training for a PINN revolves around minimizing a loss function that utilizes the governing physics PDE, along with the initial and boundary conditions of the defined physics problem. This work aims to utilize the capabilities of DNNs in the form of PINN as a tool for modeling linear and nonlinear wave equations in one of the most essential tools in ultrasound therapeutics: Focused Ultrasound (FUS) transducers. Predicting the solution of linear and nonlinear wave PDEs via PINNs enables the simulation of medical FUS Transducers using a mesh-free approach without requiring training data from a previous solution. These two main characteristics of PINNs overcome issues typically encountered in traditional numerical modeling methods and, in turn, eliminate problems such as the CoD and other related computational complexities. Our work focuses on modeling the propagation of linear and nonlinear ultrasound waves from single- and multi-element transducers using PINNs. The predicted solution is compared to a synthetic ground-truth solution produced by the Finite-Difference Time-Domain (FDTD) method for accuracy and correctness insights. This effort is steered towards studying and developing FUS transducers in contribution to their applications in various noninvasive medical treatments for a variety of tissue abnormalities. FUS beams have the potential to be focused in small volumes, which allows for a range of bioeffects depending on beam intensity. The ultimate goal of this work is to present an accurate and computationally efficient modeling tool for FUS transducers. This modeling tool will allow a physician to intraoperatively control the focus of FUS transducers at the target, thereby maximizing treatment efficiency and enabling a site-specific physician-centered workflow for noninvasive surgery.