Highly Accelerated 3D TOF MRA using Deep Learning Reconstruction with Raw K-space Simulation

Motivation: 3D-TOF-MRA suffers from long scan times that increase motion artifacts, limiting its resolution and clinical applicability. Although deep learning (DL) reconstruction shows promise for acceleration, large 3D-TOF-MRA k-space datasets for training are generally unavailable. Goal(s): To develop a DL-based reconstruction method that enables highly accelerated 3D-TOF-MRA using publicly available data for training. Approach: We simulated complex-valued multi-coil k-space from the IXI magnitude dataset and used a customized 3D variational network with architectural and data-handling improvements designed for 3D-TOF-MRA. Results: The proposed method demonstrated superior reconstruction results on in vivo data over existing methods, preserving most fine vessels with minimal artifacts with 8-fold acceleration. Impact: The proposed method shows promise for highly accelerating 3D-TOF-MRA due to its superior performance. Overcoming data scarcity, this approach holds the potential for advancing research and clinical applications of high-resolution whole-head 3D-TOF-MRA imaging, enhancing cerebrovascular diagnostics.