ML+X Seminar: Prof. Tamir - Computational MRI w Deep Learning

Magnetic resonance imaging (MRI) is a powerful non-invasive and non-ionizing medical imaging modality that offers superb soft tissue contrast and unprecedented views of anatomy and function. However, MRI is inherently slow and expensive. Recently, advances in computational imaging in tandem with deep learning have led to significantly shorter scan times and enabled new applications of MRI. In this talk, Prof. Tamir will discuss his lab’s work combining model-based algorithm unrolling with deep learning to optimize the MRI acquisition and reconstruction with the goals of reducing scan time and improving reconstruction robustness. He poses the MRI reconstruction problem as a blind multi-channel deconvolution directly in the frequency domain. The lab applies alternating minimization interleaved with convolutional neural networks for regularization and represents the reconstruction network as the unrolled iterative optimization. The reconstruction network is trained in a supervised setting and evaluated for the resulting image quality and robustness. Prof. Tamir will also show initial work to optimize the MRI scan parameters based on unrolled iterative optimization and discuss its connection to down-stream applications. Finally, he will discuss some of the lab's other ongoing work combining model-based imaging and deep learning. Speaker Bio Jon Tamir is an Assistant Professor in Electrical and Computer Engineering at UT Austin. He received his PhD in EECS from UC Berkeley. His research focus spans computational medical imaging, signal processing, and machine learning. He is primarily interested in applying advanced imaging and reconstruction techniques to pediatric MRI, with the goal of enabling real clinical adoption. He is a developer of the open-source BART toolbox for computational MRI.