Lu, Zhengyi; Liang, Hao; Lu, Ming; Martin, Dann C.; Hardy, Benjamin M.; Dawant, Benoît M.; Wang, Xiao; Yan, Xinqiang; Huo, Yuankai. (2025). PHASE: Personalized Head-based Automatic Simulation for Electromagnetic properties in 7T MRI. Magnetic Resonance Imaging, 124, 110532. https://doi.org/10.1016/j.mri.2025.110532
Accurate, personalized models of the human head are becoming essential for electromagnetic (EM) simulations, which are used to study how electric and magnetic fields interact with the body. These simulations are especially important for assessing safety limits, such as the Specific Absorption Rate (SAR) — the amount of EM energy absorbed by body tissues. Currently, most simulations rely on generic head models from the Virtual Population, since creating detailed, patient-specific models by hand is extremely time-consuming and resource-intensive.
To solve this problem, we developed PHASE (Personalized Head-based Automatic Simulation for EM properties) — an open-source toolbox that automatically builds high-resolution, individualized head models for EM simulations. PHASE uses both T1-weighted MRI and CT scans to identify and label 14 different tissue types in the head, creating anatomically realistic models tailored to each patient.
We tested PHASE using real data from 15 human patients and compared its results against gold-standard, semi-manual methods. The PHASE-generated models produced nearly identical results for both global SAR and localized SAR averaged over 10 grams of tissue (SAR-10g), showing strong accuracy and reliability.
Overall, PHASE offers a fast, automated, and accessible way to generate patient-specific head models for EM research and safety testing. By making the toolbox and models publicly available at https://github.com/hrlblab/PHASE, this work lays the foundation for creating large-scale, personalized datasets that can advance both clinical and engineering applications.
Fig. 1. An overview of our PHASE toolbox for constructing 3D human head models from T1w MRI and CT volumes and the validation process. Starting from raw T1w MRI and CT data, we employ trained deep learning models to generate a 14-label segmented human head volume. One segmented head with 14 labels by proposed method in axial, sagittal and coronal planes is shown. An optional evaluation process is performed to test the models generated by our method. The voxel models are meshed in spatial steps of in XFdtd (Remcom, 2025) and EM simulations are performed to calculate the field, global and the local SAR-10g.