Deep learning-based whole-brain B₁ ⁺-mapping at 7T.

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Artificial Intelligence in Medical Imaging
• Radiofrequency (RF) Field Calibration

Background:

• Accurate estimation of the transmit magnetic RF field (B1+) is crucial for quantitative MRI, especially at high field strengths like 7T.
• Subject-specific B1+ calibration is essential for parallel transmission (pTx) techniques to achieve homogenous RF excitation.
• Traditional B1+ mapping methods can be time-consuming, limiting their clinical applicability.

Purpose of the Study:

• To investigate the feasibility of using complex-valued neural networks (NNs) for rapid B1+ map estimation.
• To enable B1+ estimation from multi-slice localizer scans with varying slice orientations in the human head at 7T.
• To accelerate subject-specific B1+ calibration for pTx MRI applications.

Main Methods:

• Acquisition of channel-wise B1+ maps and multi-slice localizers (axial, sagittal, coronal) in 15 healthy subjects at 7T.
• Training of four complex-valued NN configurations using five-fold cross-validation, including one trained on all slice orientations.
• Validation of predicted B1+ maps against reference scans using quantitative metrics and in-vivo testing with dynamic kt-point pulses.

Main Results:

• Predicted B1+ maps showed high similarity to measured maps across different orientations.
• Mean relative errors in magnitude were low: 2.70±2.86% (transversal), 1.82±0.69% (sagittal), and 1.33±0.27% (coronal).
• A network trained on all orientations demonstrated robust B1+ prediction, enabling feasible in-vivo homogenous excitation.

Conclusions:

• Complex-valued NNs are feasible for estimating multi-slice B1+ maps from localizer scans at 7T.
• The proposed method accelerates B1+ calibration, enhancing the efficiency of pTx MRI.
• This approach holds promise for improving subject-specific RF field calibration in advanced MRI techniques.