Improving FLAIR SAR efficiency at 7T by adaptive tailoring of adiabatic pulse power through deep learning estimation

Abstract

Purpose The purpose of this study is to demonstrate a method for specific absorption rate (SAR) reduction for 2D T2‐FLAIR MRI sequences at 7 T by predicting the required adiabatic radiofrequency (RF) pulse power and scaling the RF amplitude in a slice‐wise fashion. Methods We used a time‐resampled frequency‐offset corrected inversion (TR‐FOCI) adiabatic pulse for spin inversion in a T2‐FLAIR sequence to improve B1+ homogeneity and calculated the pulse power required for adiabaticity slice‐by‐slice to minimize the SAR. Drawing on the implicit B1+ inhomogeneity in a standard localizer scan, we acquired 3D AutoAlign localizers and SA2RAGE B1+ maps in 28 volunteers. Then, we trained a convolutional neural network (CNN) to estimate the B1+ profile from the localizers and calculated pulse scale factors for each slice. We assessed the predicted B1+ profiles and the effect of scaled pulse amplitudes on the FLAIR inversion efficiency in oblique transverse, sagittal, and coronal orientations. Results The predicted B1+ amplitude maps matched the measured ones with a mean difference of 9.5% across all slices and participants. The slice‐by‐slice scaling of the TR‐FOCI inversion pulse was most effective in oblique transverse orientation and resulted in a 1 min and 30 s reduction in SAR induced delay time while delivering identical image quality. Conclusion We propose a SAR reduction technique based on the estimation of B1+ profiles from standard localizer scans using a CNN and show that scaling the inversion pulse power slice‐by‐slice for FLAIR sequences at 7T reduces SAR and scan time without compromising image quality.

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

“…This is in line with expectations based on previous work, in which magnitude B 1 maps were predicted with mean error of 9.5 ± 7.1%. 37 The experimental and 3D nature of data in Abbasi-Rad et al 37 may have caused the slightly higher error. We also observe similar B1 predicted correlation coefficients to those reported in Wu et al 36 (~0.99), which was retained even when our networks were cascaded multiple times.…”

“…Improvements to excitation phase were evident but modest. It has been previously acknowledged that phase changes due to motion are difficult to model, and other work on motion-related field changes, 22 including B 1 prediction, 37 often neglect phase altogether. Most of the phase-prediction error occurred at phase-wrapping boundaries.…”

“…Neural networks have previously been used to predict B 1 maps for the purpose of SAR reduction. 37 This was achieved through slice-wise pulse scaling based on a predicted 3D B 1 magnitude. The entire 9-cm axial slab could be predicted within approximately 0.8 seconds, permitting concatenation into pseudo-3D B 1 maps that could feasibly be used for pulse scaling or similar SAR management here.…”

“…Abbasi-Rad et al used a convolutional neural network to reconstruct B 1 + from a localizer scan for the purpose of SAR reduction through pulse scaling based on slice-wise B 1 + magnitude; however, B 1 + prediction quality was dependent on head position. 37 …”

“…This approach was limited to postprocessing; prediction quality deteriorated when was predicted directly from undersampled images. Abbasi‐Rad et al used a convolutional neural network to reconstruct from a localizer scan for the purpose of SAR reduction through pulse scaling based on slice‐wise magnitude; however, prediction quality was dependent on head position 37 …”