Generalize RF Pulse Design using Physics-Guided Self-Supervised Deep Learning

Motivation: Current DL-based RF-pulse-design methods rely on large amounts of data and can be challenging to implement. They can also be suboptimal in real-time applications with scanner imperfections. Goal(s): Introduce a Generalized RF-pulse-design method using Physics-guided Self-supervised learning (GPS) that integrates online adaptation for real-time compensation. Approach: GPS integrates the Bloch-equations as a physics-model into a self-supervised learning framework to guide and enforce the RF-pulse-design learning process. Low-rank adaptation is used to expedite the learning process for online adaptation. Results: 1D-selective, B1-insensitive, saturation and multidimensional RF-pulse-design is demonstrated. GPS's flexibility and versatility is further demonstrated by compensating for scanner imperfections (B0/B1+ inhomogeneity) real-time. Impact: GPS uses a case-specific strategy not requiring diverse training data and enables a more general RF-pulse-design approach. It can naturally be extended to other designs such as SMS and pTx, and incorporate advanced tissue models such as multi-pool systems.