A realistic human head phantom for electromagnetic detection of brain diseases

Acquisition of image data

Based on the imaging advantages of different image types, appropriate image data was used to reconstruct the corresponding structure. Table 1 provides the correlation between the structures to be reconstructed and the corresponding imaging data types. Both CT and MRI scans were performed using thin-layer scanning to ensure the accuracy of the reconstructed model. The CT and MRI scans had a layer resolution of 0.625 and 1 mm, respectively. The scanning was performed on a healthy 23-year-old Chinese male volunteer with no history of brain diseases or surgeries. The data were collected at the Army Characteristic Medical Center Daping Hospital. The study was approved by the ethics committee of the Chinese People’s Liberation Army Characteristic Medical Center (Medical Research and Ethics Review (2022) No. 147). The study obtained written informed consent from the study participants.

The CT scanning was performed using a 64-slice spiral CT scanner (64-slice LightSpeed; GE, Boston, MA, USA) at the Army Characteristic Medical Center Daping Hospital. The scanning range extended from the mandible to the cranial vertex. The specific scanning parameters included: a slice resolution of 0.625 mm, interslice spacing of 0.625 mm, tube voltage of 120 kV, tube current of 350 mA, and a pitch of 0.984. For the contrast-enhanced scan, an iodine contrast agent was injected through the elbow vein at a rate of 5 mL/s using a high-pressure injector, with a total volume of 50 mL. The arterial-phase scanning was conducted by the bolus tracking method, where the scan was triggered when the CT value in the region of interest (internal carotid artery) reached 150 Hounsfield units. The scanning direction was from the mandible to the cranial vertex. After completing the arterial-phase scanning, a 2 s interval was maintained, and then the venous-phase scanning was performed from the cranial vertex to the mandible. A total of 915 cross-sectional images were obtained for the arterial phase and 912 images were obtained for the venous phase during CTA scanning.

The MRI scans were conducted using a Siemens Verio 3.0 T superconducting MRI system with an eight-channel head coil. Two types of axial scans were performed: T₁WI and T₂WI. For the T₁WI sequence, the imaging parameters were as follows: repetition time (TR) = 2,100 ms; echo time (TE) = 2.93 ms; matrix size = 256 × 256; slice thickness = 1 mm; and field of view (FOV) = 256 × 256 mm². The total scanning time was 4 min and 16 s. A total of 160 cross-sectional images were acquired for the T₁WI sequence. For the T₂WI sequence, the imaging parameters were as follows: TR = 3,200 ms; TE = 409 ms; matrix size = 256 × 256; slice thickness = 1 mm; FOV = 250 × 250 mm². The total scanning time was 9 min and 25 s. A total of 176 cross-sectional images were acquired for the T₂WI sequence.

3D reconstruction of the model

This study used four different types of imaging data for 3D reconstruction: CTA venous phase, CTA arterial phase, MRI-T₁WI, and MRI-T₂WI. Initially, these four sets of imaging data were fused and registered (adjusting CT and MRI images to the same coordinate system to ensure that the structures reconstructed from different types of images can be restored to the head model). The CTA arterial phase data served as the template sequence and was registered with the other three types of imaging data as shown in Fig. 1. Following image fusion and registration, the fused and registered images were overlaid with the template CTA venous phase images for visual inspection. The largest error was identified and measured. If the maximum error was greater than 1 mm, the fusion and registration data were considered unacceptable. The process was repeated until the error was 1 mm or less. After successful registration, the corresponding tissue structures were segmented and reconstructed based on the image types described in Table 1. Both the image data fusion and tissue structure 3D reconstruction were carried out using the 3D Slicer software by an experienced technician proficient in human tissue structure 3D reconstruction. Figure 2 shows the segmentation and reconstruction of various organizational structures. After the segmentation of each tissue structure, the reconstructed segmented images were overlaid with the template sequence images within the designated region of interest. The largest error was identified and measured. If the maximum error exceeded 1 mm, the segmentation model was rejected and required reprocessing until the error was equal to or less than 1 mm. Once the segmentation models met the criteria, they were exported as STL format files.

Establishment of electromagnetic numerical model

In the CST electromagnetic simulation software, the 3D models of various tissue structures were imported in STP format using the “Import” function. Figure 3 depicts the human head model as imported into the CST software. Different cross-sectional views clearly displayed the various tissue structures inside the brain. The head model established in this study focuses on the influence of the dielectric properties of various parts of the head on the electromagnetic field. For tissues with similar dielectric constant and conductivity or smaller structures, they can be combined and simplified. For tissues which the dielectric constant and conductivity have a significant impact on the overall dielectric property of the head, they are reconstructed separately. The skin and skull have a large area and are the outermost tissues of the head, which is very important for studying the reflection of electromagnetic waves at the head boundary. In clinical practice, the main types of cerebral hemorrhage include basal ganglia hemorrhage, thalamic hemorrhage, cerebellar hemorrhage, and brainstem hemorrhage, with most cerebral hemorrhages occurring in the basal ganglia region. Therefore, the reconstruction of cerebral blood vessels, basal ganglia, brain, cerebellum, thalamus, and brainstem is crucial for electromagnetic detection research on cerebral hemorrhage. Cerebrospinal fluid is the most conductive part of the brain structure, so its reconstruction is also very important in the head model. Figure 4 illustrates the 14 different tissue structures that constituted the head model, namely skin, facial muscles, cranial bones, cerebrospinal fluid, gray matter of the brain, white matter of the brain, basal ganglia, thalamus, brainstem, cerebellum, cerebral arteries, cerebral veins, eyeballs, and vertebrae.

Assigning the appropriate dielectric parameters to each tissue structure at different frequencies created the human head electromagnetic numerical model. Considering the relationship between the size of the human brain and the wavelength of electromagnetic waves, the frequency of microwave brain detection is usually not more than 3 GHz. In this study, the electromagnetic model covered a frequency range of 0.3–3 GHz. Figure 5 presents the dielectric constants and tangential losses of blood, gray matter, white matter, cerebellum, cerebrospinal fluid, muscles, skin, nerves, vitreous humor (eye), and cancellous bone in the frequency range of 0.3–3 GHz, with data obtained from Gabriel’s database (Gabriel, Gabriel & Corthout, 1996; Gabriel, Lau & Gabriel, 1996). The dispersion spectrum of the tissue dielectric constants and tangential losses was sampled to generate 1,001 data points, facilitating the fitting of the material dielectric dispersion. Within the simulation software, a polynomial fit was used for dispersing each tissue material, with the fitting error controlled to be within 1%. The maximum fitting error for white and gray matters was 0.9734% and 0.1212%, respectively. After assigning the dielectric parameters to the different tissue materials, each material was designated for the corresponding structure in the human head model, resulting in the 0.3–3 GHz human head electromagnetic numerical model. The frequency range of the model could be adjusted based on specific needs by assigning the appropriate dielectric parameters to each tissue structure for the desired frequency point or range.

Stroke model

Various disease models, such as strokes, can be established based on a healthy head model. A brain hemorrhage electromagnetic numerical model was created by inserting a sphere at different positions within the brain and assigning the material of the sphere as blood. Figure 6 displays four different types of brain hemorrhage electromagnetic models: thalamic hemorrhage, brainstem hemorrhage, brain parenchymal hemorrhage, and basal ganglia hemorrhage. Different locations and degrees of brain hemorrhage could be simulated by changing the position and size of the blood sphere. Similarly, the sphere could be set with ischemic properties to establish a brain ischemia model, which implies having dielectric properties lower than those of the surrounding tissues by 10% (Semenov et al., 2017).

Electromagnetic simulation

The simulation software was CST Studio Suite 2020. The dielectric properties of each organization were derived from the database by Gabriel, Lau & Gabriel (1996). Within 0.3 to 3 GHz, 1,001 points were extracted from the database at equal intervals of frequency to fit the dielectric property dispersion spectrum of the brain tissue. In the simulation, the dispersion spectrum of brain dielectric properties was fitted by polynomial, and the fitting error was controlled within 1%. The maximum fitting error was 0.9734% for white matter and 0.1212% for gray matter.

Figure 7 demonstrates a brain hemorrhage simulation experiment conducted in the CST simulation software. A 1.2 GHz antenna was placed on the side of the head model for the simulation. The patch antenna used FR4 material as the substrate, and the metal copper foil was directly printed on the substrate. To achieve unidirectional radiation characteristics, the back side of the substrate, i.e., the ground layer, was completely covered with the metallic copper foil. The working frequency was around 1.2 GHz. A solid sphere was inserted into the right hemisphere of the brain. The material of the sphere was set to represent blood, simulating a brain hemorrhage. The radius of the sphere was set to 12.9 mm, representing a hemorrhage volume of 9 mL. Figure 7B shows the coronal section of the brain hemorrhage simulation experiment, the complete structure of the brain, as well as the location and size of the simulated brain hemorrhage. Figure 8 shows the S11 curves for 0 and 9 mL hemorrhage volumes. A significant difference was observed between the two curves near the operating frequency of the antenna. The Wilcoxon signed-rank test was performed on the S11 amplitudes for the two hemorrhage volumes within the frequency range of 1.2–1.216 GHz. The difference between the S₁₁ amplitudes for the non-hemorrhage model and the 9mL hemorrhage model was statistically significant (P < 0.05), indicating that the two models had different impacts on the reflection coefficient of the antenna. Figure 9 presents the electric field and magnetic field distribution at a frequency of 1.2 GHz in the human head model. It was evident that using a high-precision human head model with a realistic anatomical structure was necessary for accurate simulation, enabling the realistic representation of actual scenarios. This approach allowed the observation and study of the electromagnetic effects of different brain tissues and the impact of the human head on electromagnetic wave propagation.

About 9 h were cost in single-channel simulation (CST Studio Suite 2020 software, computer configuration: Intel(R) Core(TM) i7-10700K CPU @ 3.80 GHz, 32GB RAM, graphics card GTX1070Ti, hard disk with 1TB capacity), because the high-precision numerical model was created based on high-resolution MRI, CT, etc. For multi-channel microwave detection system, the number of channels is usually N(N-1)/2 (N is the number of antennas), which is generally time-consuming in simulation. However, the high-precision model is the foundation, and we can obtain a low-resolution model by down-sampling to adapt to different application scenarios. Secondly, it can also merge some small tissues to reduce the fineness of the brain structure and improve the simulation efficiency.

Design and production of molds

Each brain tissue structure model was printed as a hollow model using 3D printing technology and then filled with tissue-mimicking materials. These individual components were then assembled to construct a physical human head electromagnetic model.

The design diagram of the 3D printing model is shown in Fig. 10A. The brain, cerebellum, brainstem, thalamus, and basal ganglia were all printed as hollow models that could be individually disassembled. The outer layer of the brain was a closed cavity filled with a liquid to simulate the cerebrospinal fluid. Similarly, the outermost layer of the head model was also a closed cavity filled with the appropriate material to simulate skin. All cavities that required filling with tissue-mimicking materials were perforated to ensure the efficient filling and drainage of these materials. The chosen 3D printing material was a fully transparent stereolithography resin (JS-UV-2015-T, the dielectric constant is about 3.6 ± 1) with precise and durable properties. Using a transparent material for printing facilitated the observation of the filling status of tissue-mimicking materials within the model. Figure 10B displays the 3D-printed transparent human head model, with each structure in a hollow form, and the cavities filled with appropriate tissue-mimicking materials. The corresponding plugs were also printed to seal the cavities. Of course, using this model data can also create other types of brain molds and develop corresponding brain physics models (Zhang et al., 2017; Mohammed et al., 2021; Särestöniemi et al., 2024).

Production of physical models

The tissue-mimicking materials corresponding to the specific frequency were prepared and filled into different parts of the 3D-printed head model to create the human head electromagnetic model at the desired frequency. In this study, liquid tissue-mimicking materials for blood, gray matter of the brain, white matter of the brain, cerebrospinal fluid, cerebellum, skin, and nerves were prepared using pure water, 98% glycerol, anhydrous ethanol, methanol, and NaCl. The dielectric properties of these tissue-mimicking liquids were measured using a dielectric probe (Keysight N1501A) and a vector network analyzer (Keysight E5061B). The experimental setup for measuring the dielectric properties of tissue-mimicking liquids is shown in Fig. 11. Different liquids were added to a beaker and mixed thoroughly with a glass rod during the measurements. Care was taken to avoid air bubbles during the process. The measurements were performed once it was confirmed that no air bubbles were present in the beaker. To ensure accuracy, an adequate amount of liquid was used so that the bottom surface of the probe was at least 1 cm below the liquid surface. Table 2 lists the formulations of the tissue-mimicking liquids at 1.2 GHz. Table 3 presents the dielectric constants and tangential losses of the tissues and tissue-mimicking materials at 1.2 GHz. The dielectric properties of the tissues were obtained from Gabriel’s database. The tissue-mimicking materials exhibited a maximum deviation of 3.1% from the actual tissue dielectric constants (for nerves) and a minimum deviation of 0.26% (for skin). Regarding tangential losses, the maximum deviation from the actual tissues was 16.1% (for blood) and the minimum deviation was 0.59% (for gray matter).

Different tissue simulation solutions were prepared according to the proportions in Table 2. The prepared tissue simulation solution was filled into the 3D printed model, and the filling holes were sealed with plugs and hot melt adhesive. The filled tissue structure was assembled to form a physical model of the human head at 1.2 GHz (a common frequency used in microwave brain detection studies (Persson et al., 2014; Gong et al., 2023; Abbosh et al., 2024)). Each structure was filled and installed according to the steps from inside out. First, the brainstem and cerebellar models were filled and installed. Next, the thalamus, basal ganglia, and brain models were filled and assembled. Finally, cerebrospinal fluid and skin simulation fluid were filled, as shown in Fig. 12.

For ease of distinction and observation of different tissues, dye pigments were added to the tissue-mimicking liquids (Fig. 12). In actual experiments, dye pigments are not necessary to maintain the accurate dielectric properties of the tissue-mimicking liquids. Figure 13 presents an example where a balloon is inserted into the brain model. The blood-mimicking or ischemia-mimicking liquid can be injected into the balloon through a catheter, allowing the creation of actual electromagnetic models for cerebral hemorrhage and cerebral ischemia for physical experiments.

In the established physical head model, there were still some not significant difference with the numerical model. In order to accommodate the liquid that simulates the characteristics of brain tissue dielectric property, it was necessary to use 3D printing materials to make thin walls of different tissues, this was different from the numerical model, but it was essential in physical model. In order to minimize the impact of the wall of container, we made the container wall as thin as possible, the thickness was about 1.5 mm. We analyzed the effect of the wall of container in the model by simulation. In the experiment, the wall of container was endowed with the dielectric properties of the stereolithography resin. The results were shown in the Fig. 14. By comparing with the Fig. 9, it could be seen that the wall of container had some influence on the simulation results. Therefore, the ideal next step is to print containers with materials similar to the electrical properties of human brain tissue, although this is still relatively challenge.

Electromagnetic measurement on the physical model

Based on the physical model made above (normal state, no brain disease), we conducted an electromagnetic measurement experiment, as shown in the Fig. 15. The antenna in the experiment was made according to Fig. 7 and connected to a vector network analyzer (Keysight 5061B). To avoid introducing interference, the antenna was placed on a specially 3D-printed shelf made of resin material. The measurement results of the experiment are shown in Fig. 15. By comparing the results in the Fig. 15 with the simulation results of the numerical model without intracranial hemorrhage in Fig. 8, it can be seen that the models could provide a good continuity between simulation experiments and physical experiments for human bain detection based on electromagnetic measurement.