Using U-Nets to Predict the Effects of Head Motion on Simulated Specific Absorption Rate for Ultra-High Field...

Ultrahigh-field MRI requires careful management of the specific absorption rate (SAR), which is subject and subject-position dependent. Within-scan subject motion may exacerbate local SAR exposure, necessitating large safety margins to prevent SAR underestimation, which hampers imaging performance. This study proposes a U-Net architecture to adapt safety calculations to motion as it happens, to facilitate high-performance scanning without compromising safety.

Electromagnetic simulations were performed for five body models at multiple positions with an 8-channel parallel-transmit coil. Q-matrices were transformed into real-valued SAR distributions-to train U-Nets to estimate motion-induced effects on local SAR-which were then mapped back to Q-matrices. Separate U-Nets were trained for different types of body motion (rightward/leftward/anterior/posterior/yaw), which were then cascaded to predict the effect of composite (off-axis) and larger displacements on SAR. Finally, network-estimated local SAR distributions were compared with ground truth after-motion local SAR for realistic parallel-transmit pulses.

Subject motion had a statistically significant effect on local SAR, but network-estimated safety models recovered a faithful representation of the ground truth after-motion local SAR. For the investigated parallel-transmit pulses, the proposed approach reduced the safety margin from 2.14-fold to 1.3-fold and ensured more than 68% of the imaging performance could be realized while a safety model that includes all simulated subject positions would have limited scanning performance to as low as 21% of the maximum.

The proposed position-aware SAR calculation approach allows smaller safety margins, which has the potential to enable higher-performance UHF-MRI scanning without compromising safety for subjects who are unable to remain still.