1. Introduction
Low-frequency (LF) numerical dosimetry is an active research field due to some unresolved issues that still require to be addressed [
1,
2]. The evaluation of human exposure to the electromagnetic field requires an evaluation of the maximum value of the induced electric field in the human body. The use of the raw solution is, however, subject to numerical artifacts stemming from different sources [
3]. One common source of numerical artifacts is the so called stair-casing effect, due to the reconstruction of the body tissues using cubic elements (voxels) coming directly from DICOM (Digital Imaging and Communications in Medicine) images [
4]. Voxel-based meshes are intrinsically unable to match curved boundaries, as shown in
Figure 1.
Numerical artifacts cause an overestimation of the exposure and must therefore be filtered out with specific methods proposed in standards, guidelines and the literature [
5,
6,
7]. One possible solution to remove the stair-casing errors is the use of tetrahedral meshes. Over the last decades, thanks to the use of geometric modelling software, the organs and tissues of human models derived from DICOM images have been reconstructed using three-dimensional mathematical primitives (e.g., NURBS), giving the possibility to generate tetrahedral meshes that better reproduce the boundaries of curved surfaces. Human models discretized with tetrahedral meshes started to be used both in low-frequency and high-frequency numerical dosimetry (see, for example, [
8] where a procedure to optimize the specific absorption rate deposed in a patient during oncological hyperthermia treatment is presented). In [
9,
10], an anatomical model of the human body composed of tetrahedral elements and obtained from CT scans was used. The aim was to validate new methodologies based on A-
finite elements formulation to compute the induced currents into the human body due to LF magnetic fields generated, respectively, by realistic devices and completely unknown power sources. In [
11], two dual finite elements formulations to perform numerical dosimetry are presented. Two human models discretized with tetrahedral elements are used in the analysis of numerical errors related to the proposed methodology: the ZOL phantom built using the software AMIRA starting from the segmented data of the Visible Human Project
® (VHP), and the Ella phantom based on the Virtual Family, whose tetrahedral mesh was generated from the
voxel model by means of the free Matlab toolbox iso2mesh [
12]. It is interesting to note that in these papers the authors’ attention is focused on the validation of the new proposed formulations rather than the effects of tetrahedral meshes on numerical artifacts. The question of the influence of the quality of the mesh on computational results is explicitly addressed in [
13], where it is proposed to use a local a posteriori residual estimator to evaluate the error.
A comparison between tetrahedral and voxel-based meshes is the main objective of this paper. Regarding this aspect, some results are already available in the literature. In [
14], the authors validated their method on the tetrahedral human model ZOL and the numerical results are compared with some existing data evaluated on voxelized human models. The comparison showed some inconsistencies between the data obtained with the tetrahedral models and with the voxel-based ones. The authors underlined the difficulties in comparing different models and methods due to the fact that discretization and post-processing techniques play an important role in the process. However, they did not perform a detailed analysis on the nature of the numerical artifacts present in tetrahedral and voxel-based meshes and no conclusions can be drawn from their analysis. In [
15], the authors compared the use of voxel and tetrahedral meshes with the idea of eliminating the staircasing effect with tetrahedral meshes. Computations were performed using five 3D head models obtained from magnetic resonance imaging. The solution obtained with the smallest voxels (edge length of
) was taken as reference. Homogeneous and localized exposure were considered and, although tetrahedral meshes improved the representation of the tissue boundaries, numerical artifacts were registered and filtering techniques were still necessary. The authors conclude that low quality elements in the dosimetric domain are the reason for the failure of tetrahedral meshes in removing numerical artifacts.
In this paper, we analyze the problem from another point of view. A specific methodology is used to identify the source of numerical artifacts produced by tetrahedral meshes to provide additional details about their role in low-frequency numerical dosimetry. Simplified and realistic 3D human models are used in numerical simulations. In particular, simulations are performed on: (1) simple 3D models that also have an analytical solution, (2) 3D models that can be studied with a 2D equivalent simulation used as reference solution, (3) anatomical 3D models that make it possible to investigate the results in real exposure conditions. It is shown that in real exposure scenarios, similarly to the conclusions in [
15], numerical artifacts still require filtering techniques to be removed. However, taking advantage of specific exposure scenarios where the unique source of numerical artifacts is the stair-casing, it is shown that tetrahedral meshes are able to completely remove the errors related to geometrical modeling of the computational domain. Furthermore, it is shown that numerical artifacts can also occur in tetrahedra with very high quality if they belong to a boundary characterized by high contrast of the conductivity value. Therefore, in this paper we provide further clarifications about the role of tetrahedral meshes in numerical dosimetry with special attention to the artifacts originating from stair-casing approximation of curved boundaries of human models.
2. Numerical Formulation
In low frequency numerical dosimetry, the currents induced in the human body are too weak to modify the source field, hence human exposure can be assessed by means of the well-known scalar potential finite difference (SPFD) technique, where the magnetic field can be considered as the unperturbed source of the problem [
16]. This method has been employed extensively in the literature [
17,
18,
19,
20,
21,
22,
23]. The finite integration technique (FIT) using the nodal electric scalar potential as unknowns can be considered as a generalization of the SPFD to tetrahedral meshes, and it is used in this paper for both voxel and tetrahedral discretizations. Under this hypothesis, the linear system is
where
is the edge-to-node incidence matrix,
is the conductance matrix,
is the vector of nodal electric scalar potentials, and
is the vector of line integrals of the source magnetic vector potential along the mesh edges. The right-hand side of (
1) can be also evaluated knowing only the magnetic flux density (e.g., measurements) [
24]. The linear system (
1) is solved using the multigrid iterative solver AGMG [
25,
26,
27,
28]. All simulations in this paper have been carried out setting the relative tolerance of AGMG to
. This value has been verified to be sufficient to reach the convergence of the solution.
3. Tetrahedral Mesh Quality
A good mesh quality is an important prerequisite to ensure numerical results in agreement with the reference solution. The best quality mesh is achieved when it is uniformly composed by regular polygons in two-dimensional space and regular polyhedra in three-dimensional space. The mesh quality issue does not arise when voxelized-human models are used in numerical dosimetry because the meshes are discretized with regular cubic elements. On the other hand, the purpose of tetrahedral meshes is to reproduce curved boundaries. Therefore, it is common to find tetrahedra with much longer edges than others in the same mesh discretization. From a geometric point of view this guarantees a better reproduction of the object shape, however, from a computational point of view this affects the numerical accuracy.
Different metrics can be used to assess the mesh quality. In this paper, the quality
q of tetrahedral meshes is evaluated by using the Normalized Shape Ratio, as described in [
29], obtained as the ratio between the radius
r of the sphere inscribed in and the radius
R of the sphere circumscribed to the tetrahedron (see
Figure 2):
The factor 3 is used to normalize the value of q in the range . For a regular tetrahedron, q equals 1; therefore, a good-quality mesh should have most of tetrahedra with quality index close to 1.
4. Dosimetric Models
To make a fair comparison between the results obtained with voxel and tetrahedral meshes, it is important to remove possible geometric inconsistencies. For this reason, the tetrahedral mesh is first created with a commercial software setting a uniform mesh size. Then, starting from this tetrahedral mesh, a regular cubic discretization with the desired resolution ( or in this paper) is generated from the terahedral mesh. Each voxel is assigned the same tissue type of the tetrahedron containing the barycenter of the voxel. In this way, the tetrahedral and voxel-based meshes are as similar as possible.
4.1. Multilayered Sphere
The first model used in this paper is a multilayered sphere. The advantage of this model is that it is possible to define an analytical solution taken as reference to compare the numerical results.
Figure 3 shows the voxel-based mesh and the tissue type assigned to each layer.
Table 1 reports the details about the geometry and the tissue properties.
4.2. Human Head—Anatomical 3D Model
Voxel and tetrahedral discretizations are compared considering a more challenging model, the realistic human adult head
Colin27 Average Brain (also known as Average Colin). Average Colin is an adult brain atlas [
30] that consists of four tissues: skull, cerebrospinal fluid (CSF), gray matter (GM) and white matter (WM). This atlas was obtained by scanning the brain of Colin J. Holmes 27 times (hence the name Colin27) over the course of a few months. The images were combined to create an average brain with high structure definition. The tetrahedral mesh was created by Qianqian Fang (more information can be found in [
31]) and it is freely available for download [
32]. The model is shown in
Figure 4a.
The following tissue conductivities are used: skull , cerebrospinal fluid , grey matter and white matter . The operating frequency is .
4.3. Human Head—Simplified 3D Models
To highlight the source of numerical artifacts, before considering the complete 3D model described in
Section 4.2, two simplified 3D models were created. The first model is obtained by rotating a cutplane of Colin27 average brain (
Figure 4b) along the longitudinal axis (axis 2 of the figure) by
. The result is the axisymmetric model represented in
Figure 4c. In this case, a uniform vertical magnetic field is used as the source. The second model is obtained by mirroring the cutplane of
Figure 4b with respect to the sagittal plane and extruding the result by a thickness of
along the
z-axis. The result is the 3D representation of a planar geometry reported in
Figure 4d. An impressed external field parallel to the extrusion direction is considered as source. The added value of these simplified models is that they can be simulated as 2D axisymmetric and 2D planar problems. The 2D solutions obtained with a very fine mesh are used as reference for the 3D simulations, making it possible to investigate and understand the causes of the numerical artifacts. The mesh related to the 2D head models has been manually tuned in order to achieve accurate results without the need for any filtering of the raw solution. For instance, all sharp edges were pre-processed with an automated program, imposing a minimum curvature radius of
between each of the adjacent boundary edges, as shown in
Figure 5.
6. Conclusions
Several exposure scenarios were analyzed in order to study whether tetrahedral meshes were capable of suppressing numerical errors caused by stair-casing approximation errors in voxelized models when curved boundaries are approximated with voxels. From the analysis, we discovered that both voxelized and tetrahedral head models suffered from artifacts in the evaluation of high electric field values. In [
15], a similar conclusion was obtained but the cause of the artifacts was mainly associated with the presence of low quality tetrahedra.
In this paper, we showed that a tetrahedral mesh is indeed able to remove the source of computational artifacts related to the geometrical modeling of the computational domain (stair-casing), however, in a real exposure scenario, other sources of numerical artifacts are still present. These numerical artifacts are related to two fundamental factors: tetrahedral mesh quality and tissue contrast effect. It was shown that the conductivity contrast between neighboring tissues can cause very high electric fields even in tetrahedra with good quality index. For this reason, even when the mesh quality is close to ideal, it is almost impossible to avoid the crossing of induced currents across tissue interfaces in real exposure scenarios. As a consequence, the use of filtering techniques cannot be completely avoided.