Physics-guided self-supervised learning: Demonstration for generalized RF pulse design.

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Medical Physics
• Computational Imaging

Background:

• Designing radiofrequency (RF) pulses for MRI traditionally requires specialized algorithms for each pulse type.
• Existing methods can be complex and may not easily adapt to experimental variations or system imperfections.

Purpose of the Study:

• To introduce a generalized method for RF pulse design using physics-guided self-supervised learning (GPS).
• To leverage the Bloch equations as a guiding physics model within a self-supervised learning framework for RF pulse generation.

Main Methods:

• The GPS framework integrates a neural network with a Bloch simulator (physics module).
• The neural network learns to design RF pulses by optimizing against target profiles supervised by the Bloch simulator.
• Applied GPS to design various RF pulses, including selective, adiabatic, saturation, and multidimensional pulses, and demonstrated online adaptation for correcting system imperfections.

Main Results:

• GPS successfully designed diverse RF pulses with profiles matching target inputs in simulations and experiments.
• GPS-designed pulses exhibited unique characteristics, such as novel mechanisms for adiabaticity and reduced peak power.
• Experiments confirmed GPS's utility for online adaptation to correct imperfections like B1 inhomogeneity.

Conclusions:

• The GPS method demonstrates generalizability and flexibility for RF pulse design across multiple applications.
• This physics-guided approach offers a versatile alternative to conventional, dedicated RF pulse design algorithms.
• GPS shows significant potential for improving MRI performance through adaptive and efficient RF pulse generation.