Learning a B₀ Shimming Model Using Deep Neural Networks.

Area of Science:

• Biomedical Engineering
• Medical Imaging Physics

Background:

• Magnetic Resonance (MR) imaging relies on strong, homogeneous magnetic fields for optimal data quality and acquisition speed.
• Field inhomogeneities, caused by tissue susceptibility and hardware issues, necessitate B0 shimming to correct magnetic field variations.

Purpose of the Study:

• To develop and evaluate a Deep Neural Network (DNN) model for estimating optimal B0 shim coil coefficients.
• To assess the DNN model's performance in compensating for magnetic field perturbations under various conditions.

Main Methods:

• A simulation dataset was used to train a DNN model to predict shim coil coefficients.
• The DNN model was designed to learn complex field perturbation patterns and implicit representations of shim fields.
• Model performance was evaluated using R-squared (R²) metrics in both ideal and non-ideal shim conditions.

Main Results:

• The DNN-based shim model achieved a high performance of R²=0.941±0.005.
• The model demonstrated the ability to predict near-optimal coefficients for diverse shim volume masks.
• The DNN approach proved effective in ideal and non-ideal magnetic field scenarios.

Conclusions:

• Deep neural networks offer a fast and effective method for estimating B0 shim coefficients in MR imaging.
• This AI-driven approach has the potential to significantly reduce scan times and improve image quality in clinical MR applications.
• The proposed model shows promise for real-time application directly on the scanner.