Machine learning RF shimming: Prediction by iteratively projected ridge regression.

PURPOSE: To obviate online slice-by-slice RF shim optimization and reduce B1+ mapping requirements for patient-specific RF shimming in high-field magnetic resonance imaging. THEORY AND METHODS: RF Shim Prediction by Iteratively Projected Ridge Regression (PIPRR) predicts patient-specific, SAR-efficient RF shims with a machine learning approach that merges learning with training shim design. To evaluate it, a set of B1+ maps was simulated for 100 human heads for a 24-element coil at 7T. Features were derived from tissue masks and the DC Fourier coefficients of the coils' B1+ maps in each slice, which were used for kernelized ridge regression prediction of SAR-efficient RF shim weights. Predicted shims were compared to directly designed shims, circularly polarized mode, and nearest-neighbor shims predicted using the same features.
RESULTS: PIPRR predictions had 87% and 13% lower B1+ coefficients of variation compared to circularly polarized mode and nearest-neighbor shims, respectively, and achieved homogeneity and SAR similar to that of directly designed shims. Predictions were calculated in 4.92 ms on average.
CONCLUSION: PIPRR predicted uniform, SAR-efficient RF shims, and could save a large amount of B1+ mapping and computation time in RF-shimmed ultra-high field magnetic resonance imaging.