1. Introduction
Radioembolization (RE) is a clinical therapy for the treatment of primary or secondary hepatic tumors. The procedure is based on the administration of
90Y-loaded microspheres through the hepatic artery, which is the major blood supplier for liver malignancies [
1]. The vascular targeting makes RE extremely selective: high dose can be delivered to neoplastic areas while preserving nearby healthy tissue. Several approaches, empirical and dosimetric, have been proposed to establish the activity to be administered [
2]. Due to their simplicity, empirical models have been applied for several years, but at present they are not considered as adequate for patient-specific treatments. In fact, the clinical benefit evidence of dosimetry-based approaches [
3,
4,
5,
6,
7,
8,
9,
10] has led the scientific community to recognize the importance of accurate absorbed dose evaluation and to focus on the absorbed dose–biological effectiveness relationship.
Different dosimetric approaches for RE have been applied throughout the years: the MIRD scheme at the organ level [
11], the partition model [
12], the local energy deposition method (LDM) [
13,
14,
15], convolution calculations by voxel S-values [
16,
17,
18,
19], and direct Monte Carlo (MC) simulations [
20,
21]. Methods based on the MIRD approach at the organ level are mostly widespread due to their ease of use, but they present two main limitations: the regions of interest are considered to have homogeneous density, with uniform activity distributions. These unrealistic hypotheses can be overcome introducing the image-based 3D voxel dosimetry. Predictive patient-specific dosimetry is derived simulating the therapeutical procedure with the injection of
99mTc-macroaggregated albumin particles (
99mTc-MAA) to mimic the
90Y-microspheres distribution inside the liver. The invasive procedure and the use of nonidentical particles may affect the perfect correspondence of pre vs. post-therapeutical activity distribution, depending on patient specific vascularity, type of disease, experience, and shrewdness. However, many authors [
6,
9,
22,
23] have shown the good representativity of
99mTc-MAA activity distribution for treatment planning, which is verified afterwards by post
90Y-PET or
90Y-SPECT imaging. 3D approaches based on LDM or convolution of S-voxel values allow to account for the nonuniform activity distribution derived from the SPECT images, whereas they still assume a homogeneous density. MC simulation, unlike the previous methods, takes into account both tissue inhomogeneities and nonuniform activity distribution and is considered as the gold standard.
In the last two decades, various Monte Carlo-based internal dosimetry programs or routines have been applied [
24,
25,
26,
27,
28,
29]. Despite being considered the gold standard, they all differ in complexity and have not been integrated in the clinical practice yet, except in few centers for research purposes only. The main reason is that they usually require very high computer performances and calculation time, which are not compatible with the daily clinical routine. A second reason is often the lack of validation processes.
This work deals with the physical validation of a novel treatment planning system (TPS) named MCID (MC Internal Dosimetry tool) [
30,
31], performing Monte Carlo-based voxel dosimetry, applied to
90Y-radioembolization of liver malignancies. Patient’s CT and SPECT can be imported in MCID software, which creates different macros, i.e., sequences of scripted commands, for the simulation with GATE/Geant4 [
32,
33]. Each macro contains various settings about a specific aspect of the simulation (e.g., geometry of the system, source type, physics of the simulation etc. [
33]). MCID prepares the macros using some settings defined by the user through the platform (e.g., numbers of primaries, type of radionuclide etc.) and some default settings, such as the definition of the physics of the simulation. Most importantly, the activity source of the simulation is defined through the loaded SPECT of the patient, while the geometry of the simulation is defined after the segmentation of the loaded patient CT performed by the user on the platform (i.e., the morphological image is converted into a density map). Default and user-defined settings can be eventually modified, if necessary, manually changing the macros. These features allow the preparation of a personalized simulation, accounting for the specific patient morphology and activity distribution, in a very short time, avoiding manual coding related difficulties. An additional aim of this work was also to investigate the impact of tissue inhomogeneities on the dosimetric evaluation for the RE treatment and the potential improvement of a MC approach in this therapy.
3. Results
The computational time for each MC simulation was around 5 h. The relative statistical uncertainty on the absorbed dose value in a single voxel was below 1% for the range 150 Gy–700 Gy, below 2% for the range 50 Gy–150 Gy and below 10% for the range 1 Gy–10 Gy.
3.1. UL Case
The results for the mean absorbed dose calculation with the MIRD approach at the organ level and direct MC simulation are reported in
Table 4.
The relative statistical uncertainty on Dgate, assessed using (4), was 2% while the relative statistical uncertainty on Dmird, calculated from the uncertainty on the factor 49.38 in (2) (§ 2.2.1), was 0.1%. Therefore, the mean absorbed dose values to the liver obtained from the two methods are compatible.
3.2. SR Case
The comparison between average absorbed dose values for each sphere is reported in
Table 5.
The relative statistical uncertainties on the mean absorbed dose from GATE simulations for BS, MS and SS are 0.5%, 0.6% and 0.8%, respectively. Statistical uncertainties for OLINDA/EXM S-factors were not available.
In this case, differences are more evident, above all for SS, whose result is probably affected by partial volume effects (PVEs). In order to verify this assumption and reduce these effects, the initial activity map created with ITK-SNAP was used as input for MC simulation, skipping SPECT simulation and reconstruction. The obtained results are reported in
Table 6.
The relative statistical uncertainties on the mean absorbed dose in the three spheres obtained from the GATE simulations was 1%.
3.3. NUL Case
Several absorbed dose profiles for each subcase were extracted from different transversal slices. One example profile for the NUL-a case, presenting homogeneous liver and activity in three spherical regions and liver in concentration ratio 5:1, respectively, is reported in
Figure 5. An image of voxel-by-voxel relative difference (RD) for the same slice from which the profile was selected is also reported in
Figure 5.
All the profiles selected for the NUL-a case showed a relative difference within 3% between the absorbed dose images calculated by MC simulation and convolution of voxel S-values. Relative differences for the entire liver confirmed that this result is valid for all liver slices, except for some boundary voxels, actually external to the liver and characterized by low dose values (less than few Grays).
Absorbed dose profiles for the NUL-b case, presenting nonhomogeneous liver and nonuniform activity, are reported in
Figure 6 and
Figure 7.
The relative differences between the absorbed dose images were up to 14% in the spherical regions, having a different density (1.200 g/cm3) as compared to the surrounding liver (1.050 g/cm3).
4. Discussion
The validation of MCID platform is demonstrated at both organ and voxel level. In particular, for the UL scenario, the comparison between mean absorbed doses to liver assessed with the MIRD approach at the organ level and with the MC-based TPS showed a very good agreement (RD = 0.27%). This result highlights the efficiency of the developed dosimetric routine: under the hypothesis of homogeneous density and uniform activity distribution in each tissue, both methods are equivalent, as expected. It is interesting to point out that D
mird is a merely theoretical quantity, while D
gate depends on the image quality, e.g., partial volume effects (PVEs), which in this first case appear negligible due to the big size of the observed object (i.e., the whole liver). The effects due to image blurring become relevant instead when dealing with smaller objects, as in the SR case. This scenario allowed a comparison between the average absorbed doses to each sphere. While the BS and the MS present a RD < 8% (absolute value) between the two methods, the SS shows a dramatic RD of −69.2%, caused by the PVEs affecting the SPECT simulation with SIMIND. One contribution to PVEs rises from the matrix resampling (512 × 512 to 128 × 128), an additional contribution derives from the impulsive response of the imaging system, in the SIMIND simulation: spill-out effects affecting voxels cause a change in the activity quantified by imaging, only partially recovered with CDR corrections during the tomographic reconstruction. D
gate depends on the activity estimated from the image, while D
mird is only related to the theoretical initial activity value. The drastic impact of PVEs is evidenced by the results obtained without the SIMIND simulation: for the BS and the MS, RD reduced to 2% and 3% (absolute values), respectively, while for the SS, RD reduced to 9% (absolute value). Remaining discrepancies could be related to some differences between the MC codes used for calculating the OLINDA-EXM sphere S-factors and the updated code used in this work, and to possible mismatching in source description. Relative differences for the BS and the MS are in agreement with the results shown in [
41], where the authors presented an analogous comparison for various sphere diameters: their smallest sphere had a 27 mm diameter, resulting in RD = −5%, so the RD here obtained for SS (20 mm diameter, see
Table 6) also appears reasonable and comparable. Dosimetry in lesions with size equal (or less) to the maximum
90Y β
- range (around 12 mm in soft tissue) should be treated carefully, also due to limited resolving power of SPECT imaging. An analysis of dosimetry in small lesions and a correction factor for the MIRD standard equation are proposed in [
42].
A validation of the TPS at the voxel level was presented in the NUL-a case. The differences between the two methods (MC simulation and S-voxel convolution) for each absorbed dose profile were always within 3%.
Finally, the NUL-b scenario shows the importance of a personalized dosimetry in heterogeneous tissues, accounting for both activity distribution and density inhomogeneities, which are often overlooked in present internal dosimetry evaluations. The analysis of different absorbed dose profiles reveals that in the tumoral spherical regions absorbed dose values derived from the MC simulation are lower (up to 14% in absolute value) than absorbed dose values calculated with the convolution method, due to a density increase in lesions of about 20% with respect to the surrounding tissue. Therefore, it is necessary to include morphological patient-specific information in the treatment planning system, including careful calibration of the CT and possibly high quality CT systems (to allow HU-based density estimation for each voxel) to avoid inaccuracies based on the assumption of homogeneous tissues. The use of the highest possible quality CT could be of special value to improve the information especially in inhomogeneous tumors or e.g., in bone metastases.
The concept of voxel dosimetry is bright and suitable, but it is a highly complex association between image reconstruction, segmentation, density, PVE, activity recovery, and absorbed dose calculation. All these issues can concur to limitations which still need to be well understood and solved.
The scientific community is investing appreciable efforts to improve and assess the reliability and accuracy of dosimetry at the voxel level in volumes of interest of various scenarios.
Concerning image quantification, relevant studies are ongoing to highlight and compensate imaging and reconstruction deficiencies that may lead to unrealistic absorbed dose distributions for different organ substructures, lesions or voxel dimensions [
43].
As regards the potential impact of density and inhomogeneities, the results of this study, although theoretical, based on a MC approach represent a proof of concept and challenge analysis of more complex situations with real patient data of different clinical situations. This is in fact the topic of a current study of some of the authors, and preliminary results will be soon presented.