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

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Medical Physics
• Machine Learning

Background:

• Patient-specific radiofrequency (RF) shimming in high-field MRI requires extensive B1+ mapping and optimization.
• Current methods are computationally intensive and time-consuming, limiting clinical applicability.

Purpose of the Study:

• To develop and evaluate a machine learning approach, RF Shim Prediction by Iteratively Projected Ridge Regression (PIPRR), for predicting patient-specific RF shims.
• To reduce the need for online slice-by-slice RF shim optimization and B1+ mapping in high-field MRI.

Main Methods:

• PIPRR was developed using a machine learning approach merging learning with training shim design.
• B1+ maps for 100 human heads at 7T were simulated for a 24-element coil.
• Features derived from tissue masks and DC Fourier coefficients of B1+ maps were used for kernelized ridge regression to predict SAR-efficient RF shim weights.

Main Results:

• PIPRR predictions demonstrated 87% and 13% lower B1+ coefficients of variation compared to circularly polarized and nearest-neighbor shims, respectively.
• Achieved RF shim homogeneity and specific absorption rate (SAR) comparable to directly designed shims.
• Predictions were computed rapidly, averaging 4.92 ms.

Conclusions:

• PIPRR effectively predicts uniform and SAR-efficient RF shims for patient-specific applications.
• This method offers substantial savings in B1+ mapping and computation time for RF-shimmed ultra-high field MRI.
• PIPRR holds promise for improving the efficiency and accessibility of advanced MRI techniques.