ADVANCING IMAGE RECONSTRUCTION AND RESTORATION THROUGH ROBUST SUPERVISED AND GENERATIVE MODELS

Abstract: Magnetic Resonance Imaging (MRI) is a vital tool in medical diagnosis and treatment planning, largely due to its superior soft tissue contrast and non-ionizing imaging capabilities. However, MRI also presents challenges, such as prolonged scan times and data acquisition limitations stemming from patient privacy concerns and the heterogeneous nature of medical data. This seminar outlines a comprehensive approach that employs computationally efficient and robust deep learning algorithms to overcome these obstacles. First, I will discuss improvements to the training-data-free neural network method, Deep Image Prior (DIP), to enhance MRI reconstruction and bridge gaps in data acquisition. Next, I will emphasize the importance of robustness and generalization in medical imaging models by introducing a diffusion model that mitigates worst-case perturbations and data variations, including mask shifts and noise. Finally, I will present my future research vision, which leverages network pruning to further improve the generalization of MRI reconstruction.

Bio: Shijun liang received his B.S. degree in Biochemistry from the University of California, Davis, CA, USA, in 2017 as well as a Ph.D. degree in the Department of Biomedical Engineering at Michigan State University, East Lansing, MI, USA, in 2025. His research focuses on machine learning and optimization techniques for solving inverse problems in imaging. Specifically, he is interested in machine learning based image reconstruction and in enhancing the robustness of learning-based reconstruction algorithms

*** This Event will take place in a hybrid format. The location for in-person attendance will be room 1680 IOE. Attendance will also be available via Zoom.

Join Zoom Meeting: https://umich.zoom.us/j/93679028340

Meeting ID: 936 7902 8340

Passcode: XXXXXX (Will be sent via e-mail to attendees)

Zoom Passcode information is also available upon request to Kristi Rieger([email protected])