University Events

Description: Topic: Leveraging AI and Physics-based Models for Solving MRI Inverse Problems

Location: Zoom
https://umassd.zoom.us/j/93562746288?pwd=K2mJgnGVyxrKODTHfdet26mPWd6H5v.1
Meeting ID: 935 6274 6288
Passcode: 860060

Abstract
In the forward model of magnetic resonance imaging (MRI), the physical restrictions such as data undersampling, motion corruption, and inaccurate estimation of coil profiles cause aliasing and motion artifacts, low signal-to-noise ratio (SNR), and blurring effects in reconstructed image. State-of-the-art (SOTA) approaches to solving MRI inverse problems combine artificial intelligence (AI) and physics-based models, but they still have some limitations. The limitations include: (1) Aliasing and motion artifacts are intertwined and they are difficult to be separated and suppressed; (2) Pre-trained priors are ineffective for joint estimation of coil sensitivity and the reconstructed image; (3) Out-of-distribution problem arises in training MRI data; (4) Planning has not been explored in the context of solving the MRI inverse model; (5) There is a lack of a general prior to address multiple degradation factors in the forward model.
This doctoral proposal presents six key contributions. First, an untrained neural networks (UNN) model has been proposed for Periodically Rotated Overlapping ParallEL Lines with Enhanced Reconstruction (PROPELLER) MRI reconstruction to suppress blurring and aliasing, which incorporates physical priors. Building on this UNN, a new attention-based architecture comprising of spatial and channel attention for UNN has been developed to accelerate MRI reconstruction. Second, a novel synthetic blade augmentation technique is applied for the first time in a deep unrolled network for enhanced PROPELLER MRI reconstruction, which also introduces a novel synthetic blade generation process. Third, an ensemble-based approach is proposed to address multiple types of motion artifacts in MRI by employing ensemble of three distinct Cycle Consistent Generative Adversarial Networks (CycleGANs). Fourth, two novel priors are introduced to improve the joint sensitivity encoding (JSENSE) approach by incorporating accurate coil profile estimation within an iterative optimization framework. Building on the limitations of the two priors, a general unified prior, which is based on ensemble framework is proposed for joint sensitivity encoding to address multiple degradation factors. Fifth, AI based planning, traditionally not considered in this field, has been preliminarily studied and will be further explored in the dissertation. Finally, instead of using a specialized prior for MRI reconstruction, a general prior will be investigated to solve multiple degradation factors in the inverse model. The proposed methods will be validated through comparisons with SOTA approaches and qualitative assessments by MRI physicists. It is anticipated that these methods will advance MRI inverse problem solving and enhance MRI applications in clinical settings.

Advisor: Dr. Yuchou Chang, Department of Computer and Information Science
Committee Members: Dr. Haiping Xu, Department of Computer and Information Science
Dr. Long Jiao, Department of Computer and Information Science
Dr. Donghui Yan, Department of Mathematics
For further information please contact Dr Yuchou Chang at ychang1@umassd.edu