Project Summary
My overarching goal is to become a biomedical engineer capable of advancing minimally invasive
procedures through medical imaging and machine learning. To achieve this, I aim to develop a broad
understanding of diagnostic and interventional radiology and robust technical skills in acquiring, reconstructing,
and applying machine learning techniques to biomedical imaging data. This research focuses on addressing
unsolved engineering problems in magnetic resonance imaging (MRI)-based evaluation of focused ultrasound
for breast cancer treatment, providing a strong foundation for a successful career in interventional imaging.
Traditionally, MRI has provided qualitative insights into biological tissues. Recent advances in image
acquisition, reconstruction, and deep learning have created new opportunities for making quantitative
measurements of physical and chemical properties using MRI. Deep learning techniques face technical
challenges in quantitative MRI, such as the absence of large training datasets and their current inability to cope
with variations across scanners and protocols, especially in the interventional context. This work investigates
integrating physics knowledge into the architecture and training of deep learning models to mitigate these
problems and enable reliable and clinically deployable quantitative MRI techniques for evaluating MR-guided
focused ultrasound breast cancer treatments. Aim 1 uses physics-informed machine learning to develop an
efficient technique for measuring MR relaxation times in the breast using configuration state imaging. A
physics-informed neural network architecture and training paradigm will be methodically investigated using
simulated data. The developed model will then be trained on sparse real data acquired using a multi-echo
configuration state imaging sequence and rigorously evaluated on data from phantoms, healthy volunteers,
and breast cancer patients across multiple MR scanners and time points. Aim 2 aims to develop a time-efficient
technique for obtaining diffusion measurements in breast imaging with full 3D coverage, evaluating the relative
performance of conventional model-based techniques and physics-informed neural networks in estimating
diffusion parameters from the collected data. After developing the sequence and technique, diffusion
parameter maps will be compared with gold-standard measurements in standardized diffusion phantoms,
healthy volunteers, and breast cancer patients on multiple scanners. This research will advance the current
understanding of how to create generalizable machine learning models for MRI and to design them for usability
in a clinical context. Additionally, the developed MRI techniques will enable clinical and interventional use of
quantitative MRI, supporting the development of biomarkers that provide real-time evaluation of MR-guided
focused ultrasound breast cancer treatments.