Fri, Sep 16, 2022 @ 11:00 AM - 12:00 PM
Ming Hsieh Department of Electrical and Computer Engineering
Conferences, Lectures, & Seminars
Speaker: Yunsong Liu, Electrical and Computer Engineering, University of Southern California
Talk Title: Optimization Methods and Algorithms for Constrained Magnetic Resonance Imaging
Series: Medical Imaging Seminar Series
Abstract: Constrained MRI methods have shown great potential to improve the well-known trade-offs that exist in MRI between data acquisition time, signal-to-noise-ratio, and spatial resolution. In constrained MRI, we utilize prior information about the characteristics of the underlying MRI images to perform data acquisition, image reconstruction, and image analysis tasks more efficiently. This approach generally requires the use of mathematical optimization techniques, although the optimization problems are often challenging to solve efficiently due to their large-scale and non-trivial structure.
In this presentation, I will discuss three novel contributions I have made to mathematical optimization for constrained MRI. First, I will discuss work that utilizes phase constraints to accelerate MRI data acquisition based on non-Fourier radiofrequency encoding. While phase constraints are used classically in MRI, we believe that this is the first time that phase constraints are being applied to enable acceleration along a non-Fourier encoded spatial dimension. We make the novel observation that phase constraints can indeed be successfully used to reduce the number of required non-Fourier encodings, although this requires careful design of the non-Fourier encoding scheme. Results are presented in the context of gSlider, an acquisition method designed for highly-efficient high-resolution diffusion MRI. Second, we will describe a new algorithm we have developed that is designed for the separate regularization of magnitude and phase in MRI reconstruction problems. Our approach is based on a novel application of the proximal alternating linearized minimization algorithm (PALM), and incorporates additional novel features (i.e., Nesterov\'s momentum and independent selection of the step sizes for each coordinate) to increase convergence speed. Depending on the application, our proposed algorithm can be hundreds of times faster than existing algorithms for this problem. Finally, we will describe a novel algorithm that we have developed for spatial-spectral partial volume compartment mapping with applications to multicomponent diffusion and relaxation MRI. Our proposed algorithm is based on a novel application of the linearized alternating directions method of multipliers (LADMM) approach that takes advantage of the special structure of the inverse problem, and depending on the dataset, can achieve up to 5-fold acceleration compared to previous algorithms for this problem.
Biography: Yunsong Liu is a PhD candidate in Electrical and Computer Engineering at University of Southern California, supervised by Prof. Justin Haldar. He obtained his Bachelor\'s and Master\'s degree in Electrical Engineering at Xiamen University, China. He then spent half a year working on structured matrix recovery in Math Department at Hong Kong University of Science and Technology before joining Prof. Haldar\'s group at USC. His research has been focused on signal processing and optimization with applications in MRI.
Host: Justin Haldar, email@example.com
Audiences: Everyone Is Invited
Contact: Talyia White