Magnetic resonance imaging (MRI) is an attractive medical imaging modality as it is non-invasive and does not involve ionizing radiation. Routine clinical MRI exams obtain MR images corresponding to different soft tissue contrast by performing multiple scans. When 2D imaging is used, these scans are often repeated in other scanning planes. As a result, the number of scans comprising an MRI exam leads to prohibitively long exam times as compared to other medical imaging modalities such as CT. Many approaches have been designed to accelerate the MRI acquisition while maintaining diagnostic quality.

One approach is to collect multiple measurements while the MRI signal decays due to relaxation. This enables a reduction in scan time, as fewer acquisition windows are needed to collect the same number of measurements. However, when the temporal aspect of the acquisition is left unmodeled, artifacts appear in the reconstruction. Most often, these artifacts manifest as spatially varying image blurring. The effect is amplified when imaging in three dimensions, severely limiting scan efficiency. Volumetric variants would be used if not for the blurring, as they are able to reconstruct images at isotropic resolution.

Another established acceleration technique, called parallel imaging, takes advantage of spatially sensitive receive coil arrays to collect multiple MRI measurements in parallel. More recently, methods such as compressed sensing and its low rank extensions have been developed that leverage image structure to reduce the required number of samples for a well-posed reconstruction.

This dissertation proposes accelerated MRI acquisition and reconstruction techniques that account for the temporal dynamics of the MR signal. The methods build off of parallel imaging and compressed sensing to reduce scan time and flexibly model the temporal relaxation behavior. By randomly shuffling the sampling in the acquisition stage and imposing low rank constraints in the reconstruction stage, intrinsic physical parameters are modeled and their dynamics are recovered as multiple images of varying tissue contrast. Additionally, blurring artifacts are significantly reduced, as the temporal dynamics are accounted for in the reconstruction.

This dissertation first introduces T2 Shuffling, a volumetric technique that reduces blurring and reconstructs multiple T2-weighted image contrasts from a single acquisition. The method is integrated into a clinical hospital environment and evaluated on patients. Next, this dissertation develops a fast and distributed reconstruction for T2 Shuffling that achieves clinically relevant processing time latency. Clinical validation results are shown comparing T2 Shuffling as a single-sequence alternative to conventional pediatric knee MRI. Based off the compelling results, a fast targeted knee MRI using T2 Shuffling is implemented, enabling same-day access to MRI at one-third the cost compared to the conventional exam. To date, over 2,400 T2 Shuffling patient scans have been performed.

Continuing the theme of accelerated multi-contrast imaging, this dissertation extends the temporal signal model with T1-T2 Shuffling. Building off of T2 Shuffling, the new method additionally samples multiple points along the saturation recovery curve by varying the repetition time durations during the scan. Since the signal dynamics are governed by both T1 recovery and T2 relaxation, the reconstruction captures information about both intrinsic tissue parameters. As a result, multiple target synthetic contrast images are reconstructed, all from a single scan. Approaches for selecting the sequence parameters are provided, and the method is evaluated on in vivo brain imaging of a volunteer.

Altogether, these methods comprise the theme of MR Shuffling, and may open new pathways toward fast clinical MRI.




Download Full History