Motivation: DeepRF framework demonstrates superior capabilities in designing radiofrequency pulses using deep reinforcement learning, however its high computational demands limit its current use in clinical and industrial settings. Goal(s): Our objective is to reduce the computational time while maintaining design quality and to overcome the original framework's need for millions of pulse candidates per design. Approach: We propose FastDeepRF, a redesigned reinforcement learning architecture that observes magnetization states in real-time before deciding actions and leverages distributed training. Results: Computation time decreased from 40 hours to 2.5 hours, achieving improved pulse design quality while requiring only 10,000 pulse candidates instead of 3.8 million. Impact: FastDeepRF's dramatic reduction in computation time and increased efficiency open the way to broader adoption of AI-designed RF pulses in clinical and industrial settings, enhancing MRI exam outcomes through reduced SAR and enhanced image quality.
Datchanamourty et al. (Tue,) studied this question.