Clipped DeepControl: Deep neural network two-dimensional pulse design with an amplitude constraint layer.

Area of Science:

• Medical Imaging and Radiology
• Artificial Intelligence in Healthcare
• Computational Physics

Background:

• Advanced radio-frequency (RF) pulse design is crucial for magnetic resonance imaging (MRI) quality.
• Deep learning, including convolutional neural networks (CNNs), shows promise for accelerating RF pulse design.
• Existing CNN methods can exhibit pulse amplitude overshoots, posing a challenge despite constrained training.

Purpose of the Study:

• To enhance CNN-based RF pulse design by eliminating pulse amplitude overshoots.
• To maintain the capability of compensating for B0 and B1+ field inhomogeneities.
• To improve the safety and reliability of deep learning methods in MRI pulse design.

Main Methods:

• Implementation of a custom-made clipping layer within a convolutional neural network architecture.
• Supervised training of the CNN using optimal control pulses as constraints.
• Testing the modified CNN for its ability to predict 2D-selective RF pulses and compensate for field inhomogeneities.

Main Results:

• The CNN with the clipping layer completely eliminated pulse amplitude overshoots in the test subset.
• The method successfully preserved the ability to compensate for scan-subject dependent B0 and B1+ field inhomogeneities.
• Pulse prediction time remained significantly faster (milliseconds) compared to conventional optimal control methods.

Conclusions:

• The developed CNN with a clipping layer offers a robust and safe approach to deep learning-based RF pulse design in MRI.
• This advancement addresses a key limitation of previous CNN methods, improving reliability.
• The technique enables rapid and accurate RF pulse generation, potentially enhancing MRI efficiency and subject safety.