Physics Informed Neural Networks for Estimation of Tissue Properties from Multi-echo Configuration State MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biophysics

Background:

• Quantitative MRI techniques are crucial for precise medical diagnoses and treatment monitoring.
• Developing robust MRI methods for interventional settings requires advanced computational approaches.
• Current MRI techniques may face limitations in real-time parameter mapping during procedures.

Purpose of the Study:

• To investigate the integration of configuration state imaging and deep neural networks (DNNs) for quantitative MRI.
• To develop and evaluate DNNs for estimating magnetic resonance (MR) parameter maps from configuration state signal data.
• To establish a physics-informed framework for advancing MR parameter mapping in interventional settings.

Main Methods:

• Utilized configuration state imaging combined with deep neural networks for quantitative MRI development.
• Developed a physics modeling technique to account for inhomogeneous fields and heterogeneous tissues.
• Evaluated the theoretical capability of neural networks to estimate parameter maps using simulated configuration state signal data.

Main Results:

• Neural networks demonstrated theoretical capability in estimating parameter maps from configuration state signals.
• Different data normalization strategies showed similar performance for specific parameter estimations.
• Network architecture and data normalization significantly impacted the accuracy of estimated flip angle and T1 values.

Conclusions:

• Physics-informed machine learning, particularly DNNs, shows promise for MR parameter mapping.
• The developed signal modeling technique facilitates the creation and assessment of advanced quantitative MRI methods.
• This work supports the development of MRI techniques to guide clinical decisions during MR-guided treatments.