A physics-driven neural network with parameter embedding for generating quantitative MR maps from weighted images.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Quantitative Magnetic Resonance Imaging (qMRI)

Background:

• Traditional qMRI requires multiple scans, increasing time and limiting clinical use.
• Deep learning (DL) offers potential for synthesizing quantitative maps but often ignores MR signal physics.
• Ignoring physical principles compromises DL model performance and generalizability in qMRI.

Purpose of the Study:

• To develop a DL-based approach for accurate qMRI synthesis.
• Integrate MRI sequence parameters (TR, TE, TI) to enhance quantitative map generation.
• Improve the accuracy and generalizability of synthesized quantitative MRI from clinical weighted images.

Main Methods:

• Proposed a physics-driven neural network incorporating MRI sequence parameters (TR, TE, TI) via parameter embedding.
• The model learns the physical principles of MR signal formation.
• Input: T1-weighted, T2-weighted, T2-FLAIR images; Output: T1, T2, PD quantitative maps. Trained and evaluated on internal and external datasets.

Main Results:

• The physics-driven DL model outperformed conventional DL methods (pGAN, U-Net) across all metrics.
• Achieved low mean percentage errors (<6% for T1, <10% for T2, <5% for PD) and MAE.
• Demonstrated superior generalization by accurately generating maps for unseen pathological regions.

Conclusions:

• Embedding MRI sequence parameters enhances DL models' ability to learn MR signal physics.
• Significantly improved performance and reliability in quantitative MRI synthesis.
• This method holds potential for accelerating qMRI and increasing its clinical applicability.