A joint three-plane physics-constrained deep learning based polynomial fitting approach for MR electrical properties...

Area of Science:

• Medical Imaging
• Biophysics
• Computational Biology

Background:

• Magnetic Resonance Electrical Properties Tomography (MR-EPT) estimates in-vivo tissue electrical properties using reconstruction algorithms.
• Physics-based MR-EPT reconstructions face artifacts like noise and boundary issues.
• Deep learning (DL) MR-EPT methods are robust but require large datasets and struggle with generalization.

Purpose of the Study:

• To develop a joint three-plane, physics-constrained deep learning framework for polynomial fitting MR-EPT.
• To merge physics-based weighted polynomial fitting with DL for improved MR-EPT reconstructions.
• To enhance conductivity estimation accuracy and generalization for clinical MR-EPT applications.

Main Methods:

• A joint three-plane physics-constrained DL framework was developed, merging physics-based weighted polynomial fitting with DL.
• Deep learning models were trained on simulated brain data to predict optimal polynomial fitting weights in three orthogonal planes.
• Network weights were jointly optimized for combined conductivity reconstruction using complex B1+ data.

Main Results:

• The proposed physics-constrained DL approach improved conductivity estimation accuracy compared to single-plane methods.
• The 3D data-based method demonstrated superior performance over conventional methods in capturing anatomical detail and homogeneity.
• In-vivo application showed excellent generalization without significant errors or artifacts.

Conclusions:

• The joint three-plane physics-constrained DL framework offers improved MR-EPT conductivity estimation.
• The method enhances anatomical detail and homogeneity, outperforming conventional techniques.
• The framework's strong generalization to in-vivo data makes it suitable for clinical MR-EPT applications.