Quasi-instantaneous subject-specific slice-by-slice pTx pulse design with Deep Learning: application to 2D diffusion MRI

Motivation: Slice-specific pTx pulses can address B1+ inhomogeneity in 2D sequences at 7T however they suffer from extended calculation time making them practically unfeasible. Goal(s): Develop a deep learning method for rapid slice-by-slice pTx pulse design. Approach: An unsupervised deep learning pipeline using a ShuffleNet_V2 architecture and a novel Bloch simulation-based loss function with maximum local specific absorption rate (SAR) penalty. Results: Deep learning achieved lower Normalised-Root-Mean-Square-Error (21.56%±0.41%) than B1+-shimming (26.56%±1.10%) and the Small-Tip-Angle approximation method (25.63%±0.44%) while consuming lower specific energy deposition (SED). In-vivo evaluation in the context of diffusion MRI showed improved image quality with a calculation time of only 0.018s/slice. Impact: The proposed deep learning method enables subject and slice-specific pTx pulse design for 2D sequences with a large number of slices in a practical time, providing full brain coverage with improved homogeneity at 7T.