Parallelization and deep learning techniques for the registration and reconstruction of dynamic cardiac magnetic resonance imaging

Resumen

Magnetic Resonance Imaging (MRI) is a medical imaging technique that produces detailed images of the body organs and tissues without the use of ionizing radiation. In the field of cardiology, cardiac magnetic resonance (CMR), also known as cardiac MRI, is a valuable tool for the evaluation of patients with different pathologies. CMR modalities such as cardiac cine and first-pass perfusion allow practitioners to visualize the myocardium dynamic behavior and to evaluate blood perfusion in the heart tissue respectively, but they require breath-holds during acquisition and have technical difficulties due to the dynamics of the heart. In addition, there are other drawbacks to CMR, including the protocol duration, high costs, and the likelihood of image degradation due to irregular cardiac and respiratory motion. There is a high demand to improve the efficiency of data collection to reduce scan times or to improve spatio-temporal resolutions. To increase patient comfort, the amount of information required for image reconstruction must be reduced, i.e. collecting only a portion of k-space (undersampling), which results in increased complexity in processing the acquisitions. The issue arises because the information in the reconstruction problem is insufficient to determine a solution, i.e. the problem is ill-conditioned, so additional information and constraints must be added. This results in a cost function that is optimized to provide a regularized solution. The focus of this Thesis is on the challenge of efficiently reconstructing dynamic MRI images from highly undersampled data, taking advantage of the motion present in the dynamic sequences to utilize the redundant information in both time and space. For this purpose, we start from the groupwise compressed sensing (GWCS) solution, previously reported by our research group. On the one hand, a parallel framework for GWCS is explored. GWCS is a highly parallelizable problem, so we will make use of the OpenCLIPER framework, another previous work of ours, for its implementation. This solution is motion compensated to increase the sparse character of the solution, so we have to parallelize: 1) the groupwise registration algorithm to estimate the motion, which will be done using FFDs with cubic B-splines, and 2) the optimization algorithm of the reconstruction itself, for which we will use NESTA. Results obtained with and without motion estimation and compensation are analyzed to conclude that the solution is clinically viable in terms of execution times, and suitable for any computing device which has an OpenCL implementation. On the other hand, we propose a GWCS-like approach that leverages deep learning to enhance the reconstruction process. Our approach eliminates the need of optimization steps and utilizes deep learning techniques instead to speed up reconstructions and reduce computational complexity. We first create a fast solution for registration with unsupervised DL, called dGW, and then a self-supervised DL solution for motion-compensated reconstruction (SSMoComp) that relies on the previously trained registration. Regarding dGW, we found that it achieved comparable accuracy to traditional optimization-based approaches, but with significantly reduced registration runtimes. As for SSMoComp, we conducted a comparative analysis with a state-of-the-art solution and observed that our design outperformed it, yielding superior results. A modified version of the cine DL solution was additionally adapted for first-pass perfusion, called SECRET. Compared with state-of-the-art approaches, the SECRET method maintains good quality reconstructions for higher acceleration rates, with low training and very fast reconstruction times.