MR electrical properties mapping using vision transformers and canny edge detectors.

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Biophysics

Background:

• Magnetic Resonance (MR) imaging is a powerful non-invasive tool.
• Accurate reconstruction of electrical properties (EP) is crucial for advanced MR applications.
• Current methods for EP reconstruction face limitations in speed and accuracy.

Purpose of the Study:

• To develop a novel 3D vision transformer-based neural network for reconstructing electrical properties (EP) from MR measurements.
• To utilize MR signal characteristics and object boundaries as inputs for EP mapping.
• To establish a foundation for clinically applicable in vivo EP reconstruction.

Main Methods:

• A 3D vision transformer neural network was designed to process MR data.
• Inputs included transmit magnetic field magnitude, transceive phase, and Canny edge masks.
• The network was trained on extensive synthetic datasets and fine-tuned on realistic head models, with validation on experimental data.

Main Results:

• The model achieved high accuracy in reconstructing conductivity and permittivity, with low average peak normalized absolute errors (PNAE) below 3% for realistic head models.
• Lesion detection and characterization were accurate, with reconstructed values close to ground-truth.
• Experimental validation on phantoms and in vivo scans demonstrated the method's robustness and anatomical preservation.

Conclusions:

• A novel learning-based approach for MR-based electrical property reconstruction was successfully developed.
• The method shows significant promise for accurate and efficient in vivo EP mapping.
• This work represents a key advancement towards clinical translation of MR-based EP reconstruction.