Optimizing Radiofrequency Pulses using Deep Learning Frameworks

Motivation: Utilize modern deep learning techniques to efficiently generate RF pulses on a GPU for specific applications. Goal(s): Develop a fast, easy-to-use framework to optimize RF pulses and demonstrate its effectiveness by generating slice-selective 90° excitation and 180° plane refocusing pulses for MESE experiments. Approach: RF pulses were optimized using neural network frameworks by training to achieve a target profile on a set of simulated phantoms, in a process that mirrors neural network training. Results: The optimized pulses outperformed the SLR pulses in MESE experiments on both phantom and mouse brain. Impact: A neural network framework was developed to create high-performance RF pulses that lead to improved image quality. Constraints such as application-specific considerations and hardware limitations or perturbations can be easily incorporated into the framework for fast, easy-to-use RF pulse generation.