Using Physics-Informed Neural Networks for the prediction of the electromagnetic field in 7T Cardiac MRI

This project is a collaboration between the Comprehensive Heart Failure Center (CHFC) at the University Hospital of Würzburg and the University of Würzburg (JMU).

As a background to the project, heart failure is one of the main causes of death worldwide, and high-resolution imaging plays a very important role in diagnosing it. Cardiac MRI at 7 Tesla (ultrahigh-field) provides excellent image quality because of its high signal-to-noise ratio (SNR) and spatial resolution. 

However, its wider use is limited by the safety concerns related to the complex distribution of electromagnetic (EM) fields inside the body. These field distributions can lead to safety problems, such as localized tissue heating due to the radio frequency (RF) energy absorbed by the body during UHF MRI.

The goal of MAGNET4Cardiac7T is to create a deep learning–based method for fast, patient-specific analysis of 3D EM field distribution in the body, especially focusing on predicting Specific Absorption Rate (SAR), which is directly linked to RF heating in UHF MRI.

 

The commonly available simulations based on full-wave solutions of Maxwell’s equations are very accurate but take a long time (sometimes several days), making them unrealistic to use with patients in clinical applications. To solve this, we are developing Physics-Informed Neural Networks (PINNs). These are deep learning models that use the laws of physics (Maxwell's equations) to guide the learning process and improve the accuracy of predicting the electromagnetic (EM) field distribution inside the body.

 

From the CHFC side, we are conducting EM simulations for different RF coils (including dipoles and surface array coils) and various phantom setups. We start with simple geometries (phantoms such as spheres, cylinders, etc) and gradually move toward more complex media, such as human models. To obtain these human models, we use existing anatomical models and we also perform segmentation of CT images to generate models which can then be used in EM simulations. The results of these simulations, along with physical constraints from Maxwell’s equations, are used as training data for the deep learning models, which are being developed in collaboration with our project partner, the Data Science Department at the University of Würzburg (JMU).

References

Jabbarigargari, F., Dulny, A., Terekhov, M., Krause, A., Hotho, A., Schreiber, L. M. Prediction of EM Field Distribution in UHF MRI Using Physics-Informed Neural Networks (PINNs): Methodology and Potential Improvement in Prediction. In Proceedings of the Annual Meeting of ISMRM, Honolulu, Hawaii, USA, 2025.

Dulny, A., Jabbarigargari, F., Hotho, A., Schreiber, L. M., Terekhov, M., & Krause, A. (2025). Physical knowledge improves prediction of EM Fields. arXiv preprint arXiv:2503.11703.

Contact

Phone:  

+49 931 201-46333 (office)

+49 931 201-46300/01 (outpatient department)

 


Address:

Deutsches Zentrum für Herzinsuffizienz Würzburg | Universitätsklinikum Würzburg |
Am Schwarzenberg 15 | Haus A15 | 97078 Würzburg | Deutschland