Electromagnetic (EM) simulations have become indispensable tools in numerous medical and engineering applications [1], [2], [3] such as electromagnetic exposure assessment and radiofrequency (RF) coil design in MRI [4], [5], [6], [7], [8]. A critical component for the reliability and validity of these simulations is the use of accurate and individualized human models, which must represent precise geometric and dielectric properties of various human tissues. These models enable realistic predictions of electric and magnetic field distributions within the human body and are essential for estimating Specific Absorption Rate (SAR), thereby ensuring compliance with safety standards. This is particularly important for ultrahigh field MRI scanners with multi-channel RF transmission systems, as the EM fields and SAR distribution can vary significantly due to inter-subject variability [9], [10], [11], [12], [13], [14]. Personalized models are therefore necessary to ensure both fidelity and safety.
In this study, we introduces Personalized Head-based Automatic Simulation for Electromagnetic properties (PHASE), an automated open-source toolbox that generates high-resolution, patient-specific head models for EM simulations using paired T1-weighted (T1w) MRI and computed tomography (CT) scans with 14 tissue labels (see Fig. 1). Note that, although previous studies have demonstrated the use of MRI and/or CT scan data for generating human models [13], [14], [15], [16], this work is the first to propose and publish an open-source toolbox dedicated to this purpose. PHASE synergistically integrates T1w MRI and CT to accurately delineate and label anatomical structures. Advanced deep learning based algorithms are employed and have significantly enhanced our capability for precise and efficient tissue segmentation from MRI data. Specifically, we employed our SLANT deep whole brain segmentation model [17] to achieve fine-grained segmentation of complex brain tissues from T1w MRI data. Threshold and morphological operations are applied on CT images to obtain skull labels [18]. For comprehensive modeling, additional tissues such as fat and skin are effectively segmented using the validated DL segmentation pipelines SimNIBS [19] and GRACE [20], respectively.
To assess the performance of PHASE models (fully automated), we perform semi-automated segmentation and EM simulations on 15 real human subjects at a resolution of 1mm×1mm×1mm, serving as the gold standard reference. Each model includes detailed segmentation of 14 distinct tissue types, accurately capturing both bony structures and soft tissues. Extensive EM simulations are conducted using a realistic 16-rung high-pass birdcage coil configuration for 7 T MRI. This setup enables comprehensive analysis of SAR, including both global SAR and 10-gram averaged local SAR (SAR-10g), for quantitative evaluation. Additionally, we explore tissue groupings to evaluate how varying levels of anatomical detail affect EM simulation outcomes. The main contributions of this work are summarized as follows:
•We propose PHASE, an automatic toolbox which can automatically generate personalized human head models for EM simulation usage from pairs of real MRI and CT scans.
•We perform a comparative analysis against the outputs of manually refined reference models and existing models, evaluating the model-accuracy requirements of EM simulations.
•We explore tissue groupings to evaluate how varying levels of anatomical details affect the EM simulation results of ultrahigh field RF transmit coil.
•We release PHASE as an open resource toolbox to support research in medical physics, including applications in RF coil design and EM safety evaluation.
Comments (0)