Accelerated computational models for multi-parametric MRI reconstruction

Abstract

This thesis focuses on advancing Magnetic Resonance Spin Tomography in Time-Domain (MR-STAT), a novel quantitative MRI (qMRI) technique. Comparing to conventional contrast-weighted MR techniques which produce contrast-weighted, qualitative MR images, MR-STAT is a novel multi-parametric quantitative MR technique, which uses a single short scan and produces the quantitative images such as T1 and T2 relaxation times and proton density. These quantitative maps provide a precise and objective assessment of tissue composition and pathology. This thesis focuses on two aspects of the technical development: the optimization of the reconstruction and acquisition processes. Chapter 2 addresses this computational burden of MR-STAT reconstruction by introducing an accelerated reconstruction algorithm. This algorithm combines a neural network-based surrogate model to approximate the Bloch equation with an optimization approach using the alternating direction method of multipliers (ADMM). The new algorithm allows for a drastic reduction in computation time, enabling 2D MR-STAT reconstructions in under 3 minutes on a standard desktop PC. Chapter 3 focuses on developing a generalized surrogate model for fast MR signal simulations for transient-state gradient-spoiled sequences. The proposed Recurrent Neural Network (RNN) model has been demonstrated to be a fast and accurate model for large-scale MR signal simulations for different acquisition parameters without the need for retraining. It has been proven that the proposed RNN model can be used for large-scale simulated dataset generation and optimal sequence design, with a computation time of about 10 seconds when running on GPU cards. Chapter 4 and 5 focus on developments of novel 3D MR-STAT protocols, which could achieve high Signal-to-Noise Ratio (SNR) and high-resolution volumetric quantitative maps. In Chapter 4, a 3D MR-STAT protocol using repetitive flip-angle train and parallel imaging technique is designed and implemented. This 3D MR-STAT protocol enables the design of the two-step reconstruction algorithm: firstly, the undersampled transient-state 3D data are reconstructed by SENSE reconstruction and split into 2D MR-STAT k-space data; secondly, slice-by-slice 2D MR-STAT reconstructions can be run using the accelerated reconstruction algorithm developed in Chapter 2. A seven-minute 3D acquisition protocol is implemented and tested on in-vivo experiments of knees and lower-legs. In Chapter 5, a 3D MR-STAT protocol is optimized for whole-brain imaging. One of the main challenges in 3D MR-STAT is the presence of cerebrospinal fluid (CSF), which can cause ghosting artifacts in the reconstructed images. To address this, a time-efficient 3D MR-STAT sequence with CSF suppression is developed. This sequence is optimized to reduce the signal intensity of CSF, thereby minimizing the artifacts while maintaining high SNR in the brain's white and gray matter. The new protocol enables high-quality 1mm isotropic whole-brain relaxometry in less than 6 minutes. This method represents a major step towards the application of MR-STAT to clinical neuroimaging, where time efficiency and image quality are of great importance.