Ultrafast (milliseconds), multidimensional RF pulse design with deep learning.

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Artificial Intelligence in Medical Imaging
• Radiofrequency Pulse Design

Background:

• Advanced multidimensional radiofrequency (RF) pulses offer benefits like reduced field-of-view (FOV) imaging and improved spectroscopy.
• However, their lengthy design and computation times often limit clinical application.

Purpose of the Study:

• To develop a novel deep learning approach for the ultrafast design of multidimensional RF pulses.
• To enable real-time updates of complex RF pulses for enhanced clinical utility.

Main Methods:

• A neural network was trained to generate multidimensional RF pulses based on desired excitation regions.
• The network takes input maps of the region of interest and outputs a single-channel RF pulse.
• Training utilized a library of target RF pulses calculated via a chosen method.

Main Results:

• A simple neural network successfully generated reliable 2D spatial-selective RF pulses with performance comparable to traditional methods.
• The training library size requirement is manageable, especially for binary regions of interest.
• Numerical and experimental validation at 3 Tesla confirmed the predicted pulse performance.

Conclusions:

• Demonstrated effortless training of multidimensional RF pulses using non-MRI inputs within an MRI context.
• Achieved prediction times of milliseconds, enabling real-time updates for advanced RF pulse applications.