Digital Twins for Development of Microwave-Based Brain Tumor Detection

Brain tumors are abnormal growth of cells in the brain which can be categorized to benign (noncancerous) or malignant (cancerous) [11]. Conventionally, brain tumors are detected by magnetic resonance imaging (MRI) scanning, computed tomography (CT) scans, and positron emission tomography (PET) [12, 13]. Recently, interest in using microwave sensing/imaging technology for brain cancer detection has increased since it enables a safe, rapid, low-cost, noninvasive solution which involves nonionizing radiation, and it can be realized with portable devices [14]. The microwave technique is based on detecting differences in radio channel responses caused by abnormalities having different dielectric properties than the surrounding tissues [14]. Microwave technique has been proposed for several different brain monitoring applications such as stroke detection [15, 16], brain temperature monitoring [17, 18], and cerebral circulation monitoring [19‐21]. Additionally, its suitability for detection of brain tumors has been studied in [23‐27].

The first aim is to show how realistic digital brain tumor twins are developed from gadolium enhanced T1 weighted MRI images. The second aim is to utilize digital twins to prepare a realistic emulation and simulation environments to evaluate microwave sensing based brain tumor detection technique. The emulation environment developed with the brain tumor digital twin and realistic human tissue mimicking phantoms, i.e. replicas of human tissues in terms of size, shape, and dielectric properties. The simulation environment is created in electromagnetic simulation software platform with the developed tumor simulation model embedded with anatomically realistic human head model. The third aim is to carry out evaluations with developed emulation and simulation platforms to investigate microwave sensing technique in brain tumor detection.

This paper is organized as follows: Sect. 2 describes in detail how the brain tumor digital twins can be prepared from MRI images. Section 3 describes the realistic simulation and emulation platforms. Section 4 presents results for measurements and simulations. Finally, Conclusions and future works in given in Sect. 5.

In this section, the procedure to prepare digital brain tumor twins is explained. It includes description how to develop brain tumor models from gadolium enhanced T1 weighted MRI images and how to prepare brain tumor phantoms as well as other relevant phantoms to be used in the realistic emulation platform. Additionally, use of brain twins in realistic simulation platforms is explained as well.

Realistic-shaped and -sized brain tumor digital twin is developed by using MRI images of real brain tumor (permission obtained from the patient), which is presented using FSLeyes software image in Fig. 1. The tumor is approximately 5 cm large and located on the left temporal lobe of the brain. The aim is to convert the MRI-image to a 3D model, which can be used directly in the simulation model and/or further convert a negation of the model to obtain a tumor mold which could be used for brain tumor phantom development.

Firstly, tumor's mask is created on each MRI slice image, as shown in Fig. 2a. Tumor is cropped from each MRI-image slice manually to isolate it from the rest of the image, as shown in Fig. 2b. As a result, the tumor model shown in Fig. 2c is obtained. The model is in.nii format, which is a raster format, with files generally containing at least 3-dimensional data: voxels, or pixels with a width, height, and depth. To convert this file to smoother format and suitable for 3D printing, the second software, 3DSlicer is used.

With 3D slicer software, pixelization can be removed and unnecessary holes of the model can be filled. Figure 3 illustrates tumor model before and after smoothening with 3DSlicer. Additionally, the 3DSlicer converts the tumor model to.stl format which is suitable to be used for 3D printing or in simulation software.

Finally, Fusion360 software is used to create a mold for phantom by performing the negation of the 3D model, as shown in Fig. 4. The mold is composed of two interlocking parts including a hole which allows the insertion of the phantom mixture before solidification. Finally, the mold parts are printed with 3D printer. The mold halves are presented in Fig. 5.

Tumor phantom mixture is prepared using the recipe presented in [7] and repeated in Table 1. The aim is to obtain the same dielectric properties as measured for brain tumor in [29]. The cooking procedure is the following: First, the water is heated to 65 ℃ before gelatine is added. The oil is heated separately till 50 ℃ and added to the mixture at the same time with dishwashing liquid. The mixture is stirred smoothly and heated until 65 ℃ is achieved. The mixture is cooled slightly and poured into the mold for solidification. The solidified phantom is presented in Fig. 6.

The simulations are carried out using electromagnetic simulation software CST [32] which has several anatomically realistic human voxel models resembling humans having different sizes and body constitutions. For these evaluations, we chose the voxel model Hugo since it has the most realistic brain model. The antennas are located around the voxel model’s head similarly to a portable helmet or band type of scenario. Firstly, the simulations are carried out in the reference case, i.e. without any abnormalities. Next, the realistic tumor model is inserted into the brain model in the same location as in MRI-image hence resembling fully realistic scenario. S-parameters simulations are conducted only up to 5.8GHz to save simulation time which usually increases remarkably with voxel models with higher frequencies.

Firstly, the dielectric properties of developed phantoms are verified with SPEG’s probe. The dielectric properties of the phantoms at 2.5 GHz and 6 GHz are presented in Table 2. Also, the corresponding reference values of average human tissue, obtained from [30], are included for comparison. As stated earlier, the dielectric properties of brain tumor are obtained from [29].

It is found that the dielectric properties of the phantoms are relatively close to those of the average human tissue values. Small changes can be seen especially in conductivity values. However, as presented in [33], small changes will not impact significantly on practical scenarios, e.g., in S-parameter evaluations.

Next, the measurements are carried out with phantoms, human skull, UWB antennas and VNA. The S21 parameters, i.e. the radio channel between the antennas 1 and 2 located on the opposite sides of the head, are presented in Fig. 11. As it can be seen the tumor is clearly visible in the channel response between the antennas located on the opposite side of the sides of the skull. The impact on the channel characteristics is frequency-dependent: at certain frequency ranges the presence of tumor increases the channel attenuation and the certain ranges decrease it. A similar tendency has been observed also in the other microwave-based detection studies [16, 17]. This phenomenon is due to the several reasons: difference in dielectric properties of the tumor and brain tissue vary with the frequency and thus the diffraction caused by tumor on the propagation signal may vary clearly. Additionally, antenna radiation characteristics may vary with frequency: at certain ranges there might be a radiation null towards the tumor and at certain ranges there could be a stronger lobe towards the tumor or on the sides of the tumor. All these aspects affect together on the propagation in vicinity of tumor and hence in general inside the tissue. Hence, careful study on optimal frequency range and optimal antenna locations with realistic models is essential for microwave-based detection applications.

When comparing the reference and tumorous results obtained from the measurements, there is always a small possibility of some uncertainties, e.g., due to unintentional movements of cables, or antennas. Thus, it is good to carry out comparative simulations in which these uncertainties are removed. The simulation results for a similar setup are presented in the next section.

Finally, the simulations are carried out with CST and realistic brain tumor. The location of the tumor is set the same as with the measurement model and in MRI-images. In this case, the simulations are carried out only till 5.8 GHz to save simulation time which increases drastically as the frequency increases. Besides, at higher frequencies, the penetration of the radio signal through the whole head is not possible within due to higher propagation loss. The S-parameter results are presented in Fig. 12. It is found that also in this case, there is clear difference between the reference and tumorous cases: even up to 15 dB at 2.5 GHz. Channel attenuation decreases significantly after 3 GHz since the propagation loss in the tissue increases. Additionally. The flexible antennas are not directional and thus, less power can be directed towards the body than compared to the directive antennas. However, the advantage of the flexible antennas is excellent feasibility to the practical applications.

When comparing simulation and measurement results, it is noted that the channel attenuation is clearly stronger in the simulations than in the measurements. For instance, at 2.5 GHz, the level of S21 parameters is −60 dB in the simulations whereas only around −50 dB in the measurements. One reason for this is that the size of Hugo-voxel’s head is larger than that of a real human skull. Additionally, as noted in Fig. 10b, Hugo -voxel clearly has thicker muscle and fat layers than average in that area of the head where antennas (and tumor) are located. Especially in muscle tissue, the propagation loss is high due to high relative permittivity [30] and thus excessively thick muscle layer effects on the results remarkably. However, the trend is similar in both results: tumor causes clear differences in the channel responses. At lower frequencies, 1–1.5 GHz, the channel attenuation is stronger in the presence of tumor, whereas between 2–4 GHz it is vice versa. In the measurement the trend was somewhat similar, but the S21 curves have more fluctuation and thus there are some exceptions in this trend.

The reason for this trend is that, at lower frequencies, where the microwave signal penetrates the whole head easily, the tumor tissue with higher relative permittivity than brain tissue impedes propagation compared to the reference case. This can be seen as increased channel attenuation. Instead, at higher frequencies, where the propagation inside the tissue is more challenging due to higher loss, significant amount of the radio signal propagates as on the skin surface as creeping waves [33] or though the fat layer. In this antenna location setup case, the presence of tumor causes stronger diffractions towards on-body area, and thus the part of the in-body signal is summed to the signal travelling on the skin surface which is more prominent in this case. Consequently, the channel appears to be stronger in the presence of tumor. However, the antenna radiation characteristics at different frequencies have a clear impact on this. More comprehensive analysis of this phenomenon is left for the future work with the evaluations of several different types of UWB antennas.