Zhang, Chi2023-11-282023-11-282023-05https://hdl.handle.net/11299/258700University of Minnesota Ph.D. dissertation. May 2023. Major: Electrical/Computer Engineering. Advisors: Mehmet Akçakaya, Mingyi Hong. 1 computer file (PDF); xi, 93 pages.Accelerated MRI is widely used clinically to reduce lengthy scan times, where datasets are sampled in Fourier domain (k-space) below Nyquist rate, and recovered using reconstruction methods that utilize redundancies in datasets or measurement hardware. MRI reconstruction methods either perform interpolation in k-space, or solve an inverse problem with a known forward operator to recover the desired image. Recently, deep learning (DL) has been investigated for both types of MRI reconstruction, with improved image quality at higher acceleration rates. Among k-space interpolation methods, robust artificial-neural-networks for k-space interpolation (RAKI) has been proposed to perform nonlinear k-space interpolation, showing improved noise resilience compared to its linear predecessor (GRAPPA). However, RAKI is time-consuming since it trains multiple convolutional neural networks (CNNs) to reconstruct a single acquisition, and RAKI shows blurring at high acceleration rates, both hindering its applicability in practice. For the inverse problem type methods, physics-guided deep learning (PG-DL) has been shown to offer improved image quality and robustness than purely data-driven approaches. In PG-DL, a regularized inverse problem is solved alternatively between a proximal step solved implicitly by a neural network and a linear data fidelity (DF) step for several unrolled iterations. As a sufficient number of iterations is necessary for both unrolling and internally for DF, PG-DL leads to a deep computational graph that consumes considerable GPU memory during training. Consequently it is challenging to apply PG-DL with several acquisition types, include high-resolution, 3D/4D, and non-Cartesian MRI. Furthermore, for any DL-based approach, recent studies for single-coil datasets suggest they exhibit instabilities against adversarial perturbations, which lead no visible differences at input but impact the reconstruction results. However, the impact of adversarial perturbations against conventional multi-coil MRI remains uninvestigated, which is the more clinically used acquisition setup. This thesis studies DL-based approaches for MRI reconstruction in terms of their practical translation. First, we investigate scan-specific DL methods, starting with strategies to reduce the processing time of RAKI from hours to a clinically acceptable range within seconds. This is achieved by utilizing multi-processing and a novel line-by-line network architecture which reduces the number of CNNs trained during reconstruction. We further propose residual RAKI (rRAKI) that performs hybrid linear and nonlinear k-space interpolation, where a linear convolution provides baseline reconstruction jointly with a nonlinear component for noise and artifact reduction. Compared to linear methods, rRAKI exhibits improved noise resilience, similar to RAKI, while rRAKI offers improved image sharpness compared to RAKI. We then move to database DL approaches, focusing on PG-DL. In this setting, we tackle the memory consumption issues for large-scale PG-DL. By combining several memory-efficient techniques, PG-DL of 3D, high-resolution, multi-coil, non-Cartesian MRI is achieved. These techniques include check-pointing that keeps only one unrolled step on GPU; a Toeplitz characterization for non-Cartesian encoding and decoding operations that replace memory consuming convolutional operations with point-wise multiplications; distributed learning techniques that exploit linearity of multi-coil DF terms to distribute it into multiple GPUs; and mixed precision training that trains image-domain CNN via half-precision, and DF that works with k-space data in single precision. We then propose a 2.5D PG-DL for 3D reconstruction to tackle the issue of insufficient training data in these 3D non-Cartesian applications. The 2.5D PG-DL consists of three 2D CNNs that treat 3D volumes as batches of 2D images from three orthogonal views, such that the 2D CNNs have access to a sufficient number of 2D images from limited 3D volumes during the training. The 3D reconstruction is formulated using three constraints over 2D views and finally unrolled via variable splitting with quadratic penalty. Our results suggest that the proposed 2.5D PG-DL offer improved image sharpness compared to conventional 3D PG-DL for limited training data. Finally, we investigate the impact of adversarial attacks on conventional multi-coil MRI techniques including SENSE, GRAPPA and compressed sensing. Our results show that adversarial perturbations generated on the zero-filled image, with constraints on its l∞ norm, impact all conventional multi-coil MRI techniques, leading to substantial degradation in reconstruction at high acceleration rates. Our results also suggest that the main cause of failure in PG-DL methods against adversarial attack is the linear DF step rather than the nonlinear regularizer.enImage ReconstructionInverse ProblemMachine LearningMRIThesis or Dissertation