Fig.1. The deep reinforcement learning (DRL) pulse design framework. In a forward pass, the neural network takes the input target slice profile and outputs the predicted RF pulse. The Bloch simulator embedded in the system gives the produced slice profile, which in turn is used (along with the target slice profile) to evaluate the loss. In a backward pass, the gradient of the loss with respect to the weights of the neural network are calculated using backpropagation to update the neural network. The procedures described above are iterated during the training.

Fig.5. Comparison of the SLR algorithm vs our DRL method when used to design (A) 90◦ excitation, (B) refocusing, (C) inversion and (D) 90◦ excitation multiband (MB=2, gap=50 mm) pulses. For each scenario a separate neural network was trained. Note that despite larger fluctuation observed in the predicted RF pulse, our DRL method led to visually comparable slice profiles for refocusing and inversion, while improving slice profiles for single band and multiband excitation pulses with noticeable suppression of stopband ripples.