Bridging Mri Reconstruction Across Eras: From Novel Optimization Of Traditional Methods To Efficient Deep Learning Strategies

Abstract

Magnetic Resonance Imaging (MRI) has been extensively used as a non-invasive modality for imaging the human body. Despite substantial advances over the past decades, scan duration remains as a principal issue for MRI scanning, requiring novel techniques to accelerate data acquisition. Such techniques are poised to improve clinical patient throughput, reduce motion artifacts, enhance subject comfort, and allow higher resolution imaging in many applications. Several methods have been proposed to accelerate MRI scans. In parallel imaging (PI), k-space data was acquired at a sub-Nyquist rate with with multiple receiver coils, and the redundancy among these coils were used for image reconstruction. Following the clinical impact and success of PI methods, compressed sensing (CS) techniques were developed to reconstruct images by using compressibility of images in a pre-specified linear transform domain. Transform learning (TL) was another line of work that learned the linear transforms from data, while enforcing sparsity as in CS. Recently, deep learning (DL) has shown great promise for MRI reconstruction, especially at high acceleration rates where other traditional methods would fail. Specially, physics-guided DL (PG-DL) unrolls a traditional optimization algorithm for solving regularized least squares for a fixed number of iterations, and uses neural networks to implicitly perform regularization. These unrolled networks are trained end-to-end with large databases, using well-designed loss functions and advanced optimizers, usually using a reference fully-sampled image for supervised learning. Several approaches have noted the difficulty or impossibility of acquiring fully-sampled data in various MRI applications. Among these, self-supervised learning with data undersmapling (SSDU) was developed to allow training without fully-sampled data, and multi-mask SSDU was subsequently proposed for better reconstruction quality at high acceleration rates. Although PG-DL generally shows strong ability for excellent reconstruction performance, there are concerns for generalizabilty, interpretability and stability issues. In this thesis, we aimed to bridge the gap between traditional and DL methods, while also extending the utility of DL methods for non-Cartesian imaging. We first revisited l1-wavelet CS reconstruction for accelerated MRI by using modern data science tools similar to those used in DL for optimized performance. We showed that our proposed optimization approach improved traditional CS, and further performance boost was observed by incorporating wavelet subband processing and reweighted l1 minimization. The final version reached a performance similar to state-of-the-art PG-DL, while preserving better interpretability by solving a convex optimization problem in inference time. Second, we combined ideas from CS, TL and DL to enable the learning of deep linear convolutional transforms in a format similar to PG-DL. Our proposed method performed better than CS and TL, and gave similar performance as state-of-the-art PG-DL. It used a linear representation of image as regularization at inference time, and enabled convex sparse image reconstruction that may have better robustness, stability and generalizability properties. Third, we adapted a self-supervised PG-DL technique to non-Cartesian trajectories and showed its potential for reconstructing 10-fold accelerated spiral fMRI multi-echo acquisitions. Our proposed approach gave substantial improvements in reconstructed image quality over conventional methods. Furthermore, the blood oxygenation level dependent (BOLD) signal analysis of our proposed method provided meaningful sensitivities, with similar activation patterns and extent to the expected baselines.