Joint [Formula: see text] and Image Reconstruction in Low-Field MRI by Physics-Informed Deep-Learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biophysics

Background:

• Low-field Magnetic Resonance Imaging (MRI) offers accessible imaging in resource-limited settings and critical care.
• Image distortion and noise are significant challenges in low-field MRI, limiting diagnostic accuracy.
• Existing reconstruction methods struggle to adequately address these artifacts in low-field environments.

Purpose of the Study:

• To develop and validate a model-based image reconstruction approach using unrolled neural networks for low-field MRI.
• To correct for image distortion and noise, specifically addressing B0 field inhomogeneity.
• To enable fast and accurate image reconstruction in challenging low-field MRI scenarios.

Main Methods:

• A novel physics-informed neural network architecture, SH-Net, was developed, incorporating spherical harmonic coefficient estimation for smooth B0 field map estimation.
• The SH-Net was integrated into an end-to-end trainable model for joint estimation of the B0 field map and the MR image.
• Experiments were performed on retrospectively simulated low-field knee MRI data.

Main Results:

• The proposed physics-informed neural network approach demonstrated superior performance compared to purely model-based methods.
• Improvements in Peak Signal-to-Noise Ratio (PSNR) reached up to 11.7%.
• Root Mean Square Error (RMSE) was reduced by up to 86.3%, indicating enhanced image and field map reconstruction accuracy.

Conclusions:

• The end-to-end trained, model-based approach effectively reconstructs images and B0 field maps in the low-field MRI regime.
• This method significantly outperforms existing techniques for low-field MRI.
• The developed approach facilitates accurate low-field imaging with B0 inhomogeneity compensation across diverse conditions.