HYPR-L0: A Hybrid Technique for CE MRA with Extreme Data Undersampling Factors

Introduction. Many clinical applications, most notably contrast-enhanced (CE) time-resolved angiographic imaging, would benefit from volumetric coverage with high spatial and temporal resolution to capture the rapid passage of contrast material. For many applications such as intracranial examinations, these requirements necessitate data undersampling by as much as 40-100 times in 2D. Image reconstruction from such highly undersampled data poses a challenging problem even for recently introduced techniques created for accelerated imaging. Indeed, the acceleration from parallel MRI and k-t BLAST/SENSE [1] is limited by SNR losses and the ability to learn temporal behavior from low resolution training images. HYPR reconstruction allows much higher undersampling factors due to inclusion of a composite image that improves SNR and spatial resolution [2]. However, at extreme accelerations HYPR images may suffer from cross talk between vessels with different time courses due to very low spatial resolution and SNR of the low resolution images used in the reconstruction. We propose a synergistic approach that utilizes independent acceleration mechanisms of parallel MRI, novel compressed sensing (CS) approach, and HYPR to achieve the needed acceleration, spatial resolution and high SNR. Theory and Methods Algorithms of the HYPR family produce images by constraining the reconstruction process by a temporally averaged composite image. In HYPR LR [3], individual images are obtained by multiplying the composite image by a weighting image formed as a ratio of low resolution versions of the time frame and composite images. However, at extreme acceleration, HYPR images may suffer from spatial mixing of temporal information due to in-plane smoothing. In our algorithm, we propose to substitute the low resolution images with the corresponding images obtained from a constrained reconstruction using information about vessel edge locations in the composite image. This is accomplished by solving the following problem: ( ) 2 2 1/ 2