Rigid motion-resolved prediction using deep learning for real-time parallel-transmission pulse design.

Area of Science:

• Medical Imaging
• Magnetic Resonance Imaging (MRI)
• Artificial Intelligence in Medicine

Background:

• Parallel-transmit (pTx) pulses enable uniform excitation at 7 T MRI but are susceptible to head motion.
• Head motion introduces significant artifacts and errors in excitation profiles, compromising image quality.

Purpose of the Study:

• To develop a deep learning framework for real-time estimation of B1 field distributions during within-slice head motion.
• To enable real-time redesign of pTx pulses to counteract motion-induced artifacts.

Main Methods:

• Conditional generative adversarial networks were trained on simulated data to predict B1 maps following head displacement.
• pTx pulses were redesigned using predicted B1 maps, and their performance was evaluated against ground-truth motion-affected B1 maps.

Main Results:

• Predicted B1 maps showed high correlation with ground-truth maps, with lower error in 99% (magnitude) and 67% (phase) of evaluations.
• Redesigning pulses using predicted B1 maps reduced worst-case flip-angle normalized RMS error by 59% compared to motion-induced error.

Conclusions:

• A deep learning framework effectively predicts B1 maps online, facilitating real-time pTx pulse redesign.
• This approach mitigates head motion-related errors in pTx MRI, improving robustness and reliability.