Reconstruction Methods for Free-Breathing Dynamic Contrast-Enhanced MRI

Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) is a valuable diagnostic tool due to the combination of anatomical and physiological information it provides. However, the sequential sampling of MRI presents an inherent tradeoff between spatial and temporal resolution. Compressed Sensing (CS) methods have been applied to undersampled MRI to reconstruct full-resolution images at sub-Nyquist sampling rates. In exchange for shorter data acquisition times, CS-MRI requires more computationally intensive iterative reconstruction methods. We present several model-based image reconstruction (MBIR) methods to improve the spatial and temporal resolution of MR images and/or the computational time for multi-coil MRI reconstruction. We propose efficient variable splitting (VS) methods for support-constrained MRI reconstruction, image reconstruction and denoising with non-circulant boundary conditions, and improved temporal regularization for breast DCE-MRI. These proposed VS algorithms decouple the system model and sparsity terms of the convex optimization problem. By leveraging matrix structures in the system model and sparsifying operator, we perform alternating minimization over a list of auxiliary variables, each of which can be performed efficiently. We demonstrate the computational benefits of our proposed VS algorithms compared to similar proposed methods. We also demonstrate convergence guarantees for two proposed methods, ADMM-tridiag and ADMM-FP-tridiag. With simulation experiments, we demonstrate lower error in spatial and temporal dimensions for these VS methods compared to other object models. We also propose a method for indirect motion compensation in 5D liver DCE-MRI. 5D MRI separates temporal changes due to contrast from anatomical changes due to respiratory motion into two distinct dimensions. This work applies a pre-computed motion model to perform motion-compensated regularization across the respiratory dimension and improve the conditioning of this highly sparse 5D reconstruction problem. We demonstrate a proof of concept using a digital phantom with contrast and respiratory changes, and we show preliminary results for motion model-informed regularization on in vivo patient data.PHDElectrical Engineering: SystemsUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttps://deepblue.lib.umich.edu/bitstream/2027.42/138498/1/mtle_1.pd