Investigating Multiple Physical Priors in Deep Learning for Electrical Properties Reconstruction in MREPT

Magnetic Resonance Electrical Properties Tomography (MREPT) is a valuable imaging modality for non-invasive, in-vivo mapping of conductivity and permittivity. While deep learning (DL) based MREPT methods help mitigate the sensitivity and artifact issues of traditional approaches, their reliability and generalization remain significant challenges, limiting clinical applicability. To address these issues, this study integrates physics-informed constraints into network training, incorporating gradient similarity, Helmholtz equation constraints, complex arithmetic operations, and prior knowledge of curl and Laplacian operators. A quantitative analysis of different physical priors is conducted to assess their impact on reconstruction accuracy and model generalization. The proposed physics-informed deep learning framework demonstrates improved generalization and predictive accuracy, achieving promising results in real human head EP reconstruction. These findings suggest that multi-physical constraint can enhance the robustness of DL-based MREPT, paving the way for future clinical applications.Clinical Relevance- This study improves the accuracy and reliability of deep learning-based MREPT, enhancing its potential for non-invasive brain conductivity and permittivity mapping in clinical applications.