Research Article| Volume 97, P25-35, May 2022

Ok

# Proof-of-concept of DosiTest: A virtual multicentric clinical trial for assessing uncertainties in molecular radiotherapy dosimetry

Open AccessPublished:March 24, 2022

## Highlights

• Feasibility of setting up a ‘virtual’ multicentric clinical dosimetry trial.
• Generating reference dosimetry based on GATE Monte Carlo simulation.
• Generating centre-specific simulated patient SPECT/CT datasets with GATE.
• Comparison of reference dosimetry against the image (simulated)-based dosimetry.

## Abstract

Clinical dosimetry in molecular radiotherapy (MRT) is a multi-step procedure, prone to uncertainties at every stage of the dosimetric workflow. These are difficult to assess, especially as some are complex or even impossible to measure experimentally.
The DosiTest project was initiated to assess the variability associated with clinical dosimetry, by setting up a ‘virtual’ multicentric clinical dosimetry trial based on Monte Carlo (MC) modelling.
A reference patient model with a realistic geometry and activity input for a specific tracer is considered. Reference absorbed dose rate distribution maps are generated at various time-points from MC modelling, combining precise information on density and activity distributions (voxel wise). Then, centre-specific calibration and patient SPECT/CT datasets are modelled, on which the clinical centres can perform clinical (i.e. image-based) dosimetry. The results of this dosimetric analysis can be benchmarked against the reference dosimetry to assess the variability induced by implementing different clinical dosimetry approaches.
The feasibility of DosiTest is presented here for a clinical situation of therapeutic administration of 177Lu-DOTATATE (Lutathera®) peptide receptor radionuclide therapy (PRRT). From a real patient dataset composed of 5 SPECT/CT images and associated calibrations, we generated the reference absorbed dose rate images with GATE. Then, simulated SPECT/CT image generation based on GATE was performed, both for a calibration phantom and virtual patient images. Based on this simulated dataset, image-based dosimetry could be performed, and compared with reference dosimetry. The good agreement, between real and simulated images, and between reference and image-based dosimetry established the proof of concept of DosiTest.

## 1. Introduction

Molecular radiotherapy (MRT) is a systemic use of radiolabelled vectors (radiopharmaceuticals) for the treatment of benign and malignant tumours. A radiopharmaceutical consisting of a radionuclide linked to a targeting molecule binds specifically to tumours to selectively irradiate the tumour cells while sparing healthy tissues. In clinical practice, patients are most often administered fixed activity (under the “radioactive chemotherapy” paradigm) [
• Flux G.D.
• Sjogreen Gleisner K.
• Chiesa C.
• Lassmann M.
• Chouin N.
• Gear J.
• et al.
From fixed activities to personalized treatments in radionuclide therapy: lost in translation?.
]. However, significant variability in tumour uptake and radioactivity clearance from patient to patient can be observed. Patient-specific dosimetry allows a major paradigm shift in the administration of MRT from a “one-size-fits-all” approach to real personalised medicine where administered activity is assessed specifically for each patient.
Patient-specific dosimetry generally relies on quantitative imaging (determining the spatial distribution of the radiopharmaceutical in different organs via imaging at different times), radiopharmacokinetic assessment (determining the total number of radioactive decays in all source organs by integrating activity over time post activity administration) and absorbed dose calculation (based on radioactive decay distribution, energy emitted per decay for the radioisotope, and radiation interactions within propagating media).
The accuracy of clinical dosimetry depends on the accuracy of each step in the complete dosimetric chain. Until now, the techniques developed and employed in clinics vary in approach and sophistication, encompassing a wide range of dosimetric approaches [
• Sjögreen Gleisner K.
• Spezi E.
• Solny P.
• Gabina P.M.
• Cicone F.
• Stokke C.
• et al.
Variations in the practice of molecular radiotherapy and implementation of dosimetry: results from a European survey.
]. Strategies have been proposed to improve activity quantification and standardise dosimetric protocols [
• D’Arienzo M.
• Cazzato M.
• Cozzella M.L.
• Cox M.
• D’Andrea M.
• Fazio A.
• et al.
Gamma camera calibration and validation for quantitative SPECT imaging with 177Lu.
,
• Gregory R.
• Flux G.
• Newbold K.
• Du Y.
• Moss L.
• et al.
SELIMETRY—a multicentre I-131 dosimetry trial: a clinical perspective.
,
• Wevrett J.
• Fenwick A.
• Scuffham J.
• Johansson L.
• Gear J.
• Schlögl S.
• et al.
Inter-comparison of quantitative imaging of lutetium-177 (177Lu) in European hospitals.
,
• Zhao W.
• Esquinas P.L.
• Hou X.
• Uribe C.F.
• Gonzalez M.
• Beauregard J.-M.
• et al.
Determination of gamma camera calibration factors for quantitation of therapeutic radioisotopes.
,
• Peters S.M.B.
• van der Werf N.R.
• Segbers M.
• van Velden F.H.P.
• Wierts R.
• Blokland K.A.K.
• et al.
Towards standardization of absolute SPECT/CT quantification: a multi-center and multi-vendor phantom study.
,
• Taprogge J.
• Leek F.
• Schurrat T.
• Tran-Gia J.
• Vallot D.
• Bardiès M.
• et al.
Setting up a quantitative SPECT imaging network for a European multi-centre dosimetry study of radioiodine treatment for thyroid cancer as part of the MEDIRAD project.
,
• Frezza A.
• Desport C.
• Uribe C.
• Zhao W.
• Celler A.
• Després P.
• et al.
Comprehensive SPECT/CT system characterization and calibration for 177Lu quantitative SPECT (QSPECT) with dead-time correction. EJNMMI.
,

Dewaraja Y, Frey E, Sunderland J, Uribe C. Dosimetry Challenge https://therapy.snmmi.org/SNMMI-THERAPY/Dosimetry_Challenge.aspx (accessed October 13, 2021).

], some of them based on Monte Carlo modelling of image standards [
• Brolin G.
• Gleisner K.S.
• Ljungberg M.
Dynamic 99mTc-MAG3 renography: images for quality control obtained by combining pharmacokinetic modelling, an anthropomorphic computer phantom and Monte Carlo simulated scintillation camera imaging.
,
• Gustafsson J.
• Brolin G.
• Cox M.
• Ljungberg M.
• Johansson L.
• Gleisner K.S.
Uncertainty propagation for SPECT/CT-based renal dosimetry in 177Lu peptide receptor radionuclide therapy.
,
• Brolin G.
• Gustafsson J.
• Ljungberg M.
• Gleisner K.S.
Pharmacokinetic digital phantoms for accuracy assessment of image-based dosimetry in (177)Lu-DOTATATE peptide receptor radionuclide therapy.
].
One of the challenges to achieve that goal is to assess the accuracy of the entire chain starting from scintigraphic imaging to absorbed dose calculation, especially as the absorbed doses within the patient cannot be measured experimentally in situ. As a result, the DosiTest project [

Dositest project. http://www.dositest.org (accessed August 24, 2021).

] was initiated with the aim of evaluating the uncertainties associated with each of the steps that compose the clinical dosimetry workflow, and to propose standardised approaches to clinical dosimetry in molecular radiotherapy.
Fig. 1 depicts the schematic overview of DosiTest. The core concept is to circulate a virtual patient dataset across participating centres, who will conduct dosimetry on it according to their dosimetric protocols, just as if it were a patient enrolled in their clinical centre. The participating centres and our centre are referred to as ‘clinical centre’ and ‘host centre’ respectively in this article.
Within this multicentric clinical dosimetry trial, the reference dosimetry is obtained via Monte Carlo modelling, from the virtual patient geometry and associated reference pharmacokinetics.
Then, based on the virtual patient geometry and reference activity distribution, images are generated (via Monte Carlo modelling) and sent to each clinical centre. Absorbed doses obtained from each centre (image-based dosimetry) along with intermediary results are compared against the reference dosimetry.
The principle of DosiTest is derived from the QuantiTest project [

Hapdey S, Soret M, Ferrer L, Koulibaly PM, Henriques J, Gardin I, et al. Quantification in SPECT: myth or reality? A multi-centric study. IEEE Symposium Conference Record Nuclear Science 2004, vol. 5, 2004, p. 3170–3 Vol. 5.

]. The first presentation of DosiTest was given by Ferrer et al. [

Ferrer L, McKay E, Lisbona A, Kraeber-Bodere F, Bardies M. DosiTest: Accuracy of a radio-immunotherapy dosimetry protocol. J Nucl Med 2009;50:1865–1865.

] at the Society of Nuclear Medicine & Molecular Imaging (SNMMI) meeting in 2009. A preliminary phase of DosiTest included the design of TestDose, a tool designed to generate scintigraphic images (planar or SPECT) and to perform absorbed dose calculations [

McKay E, Ferrer L, Barbet J, Bardies M. TestDose: Software for creating dosimetry problems. J Nucl Med 2009;50:1864–1864.

,
• Garcia M.-P.
• Villoing D.
• McKay E.
• Ferrer L.
• Cremonesi M.
• Botta F.
• et al.
TestDose: A nuclear medicine software based on Monte Carlo modeling for generating gamma camera acquisitions and dosimetry.
]. Villoing et al. [
• Villoing D.
• Marcatili S.
• Garcia M.-P.
• Bardiès M.
Internal dosimetry with the Monte Carlo code GATE: validation using the ICRP/ICRU female reference computational model.
] addressed the computation of absorbed dose based on Monte Carlo modelling at the clinical scale and concluded that GATE could be safely used for radiopharmaceutical voxel-based dosimetry. However, even though simulation of complex patient dataset was technically feasible, a major limitation was the generation of realistic images that could be integrated in a clinical workstation “as if it was a real patient” (due to proprietary tags in the image header). Feasibility of 2D planar scintigraphic image generation with GATE following Lutathera® administration was illustrated by Costa et al. [
• Costa G.C.A.
• Bonifácio D.A.B.
• Sarrut D.
• Cajgfinger T.
• Bardiès M.
Optimization of GATE simulations for whole-body planar scintigraphic acquisitions using the XCAT male phantom with 177Lu-DOTATATE biokinetics in a Siemens Symbia T2.
]. We recently proposed an example of full SPECT/CT image generation in a similar clinical context [
• Kayal G.
• Chauvin M.
• Vergara-Gil A.
• Clayton N.
• Ferrer L.
• Moalosi T.
• et al.
Generation of clinical 177Lu SPECT/CT images based on Monte Carlo simulation with GATE.
].
The present work highlights the proof-of-concept of the whole DosiTest project, with an illustrated example starting from a virtual patient derived for a clinical dataset and associated activity distribution, the computation of the reference dosimetry, and the generation of 3D patient SPECT/CT images suited to image-based dosimetry. An evaluation of the goodness of the various modelling steps was performed.

## 2. Methods

### 2.1 Clinical dataset

The reference dataset was derived from the dosimetric study of a patient administered with an activity of 6.85 ± 0.34 GBq of Lutathera® (177Lu -DOTATATE), corresponding to the third therapy cycle, in the Tygerberg Hospital (South Africa), as part of an IAEA-Coordinated Research Project (CRP) E23005 on “Dosimetry in Radiopharmaceutical therapy for personalised patient treatment” [

Dosimetry in molecular radiotherapy for personalized patient treatments. https://www.iaea.org/projects/crp/e23005 (accessed September 27, 2021).

].
Patient SPECT/CT projections were acquired on a dual-headed Infinia Hawkeye 4 SPECT/CT system (GEHC, Milwaukee, USA) with a medium energy collimator and a 9.5 mm (⅜″) NaI crystal, at five time points post activity administration. Step and shoot SPECT acquisitions were performed with an energy window of 208 keV ± 10% (187.2–228.2 keV), matrix size of 128 $×$ 128 (4.418 $×$ 4.418 mm2) and 15 s per projection. The 4-slice detector array CT system rotates at 2.6 revolutions per minute helically with a fixed pitch of 1.9 mm per revolution. Other CT characteristics include a voltage of 140 kVp, tube current of 2.5 mA and slice thickness of 10 mm. CT slices were acquired in a 256 × 256 matrix with pixel size of 2.209 × 2.209 mm2.
Along with this, SPECT/CT projections of a homogeneously water-filled NEMA-IEC calibration phantom were acquired. The 177Lu activity concentrations were 32.26 kBq/ml in the background and 250 kBq/ml in the three largest spheres (22 mm, 28 mm and 37 mm).
SPECT/CT projections (both patient and phantom) were reconstructed on a Hermes™ workstation (Hermes Medical Solutions, Stockholm, Sweden) with OSEM algorithm (5 iterations, 16 subsets, 0.8 cm FWHM 3D Gaussian post filter) and corrections were applied (CT-based attenuation correction, built-in Monte Carlo-based scatter correction and default collimator-detector response). The matrix size of the reconstructed SPECT/CT images was 128 $×$ 128 $×$ 90 (4.418 $×$ 4.418 $×$ 4.418 mm3). All reconstructions discussed in this article were performed using the aforementioned software and parameters.
A calibration factor of 5.33 ± 0.27 cps/MBq was obtained from the experimental calibration images, which allowed generating activity-indexed images.

### 2.2 Patient model

A digital patient model was derived from the clinical dataset. Density and activity distributions were defined voxel by voxel, for each acquisition time point.

#### 2.2.1 Description of patient geometry

The 3D geometry map was obtained by resampling the CT to the SPECT resolution using Lanczos interpolation [

Meijering EHW, Niessen WJ, Pluim JPW, Viergever MA. Quantitative Comparison of Sinc-Approximating Kernels for Medical Image Interpolation. Medical Image Computing and Computer-Assisted Intervention – MICCAI’99, Springer Berlin Heidelberg; 1999, p. 210–7.

] and further rescaling to obtain density matrices. This was done for the CTs at all time points using OpenDose3D [
• Vergara-Gil A.
• Amato E.
• Auditore L.
• Brenet M.
• Chauvin M.
• Clayton N.
• et al.
OpenDose3D: A free, collaborative 3D Slicer module for patient-specific dosimetry.
], a clinical dosimetry module developed by our team within the 3D Slicer environment [
• Pinter C.
• Lasso A.
• Wang A.n.
• Jaffray D.
• Fichtinger G.
SlicerRT: radiation therapy research toolkit for 3D Slicer.
]. The patient geometry model therefore consisted in a matrix of 128 $×$ 128 $×$ 90 with voxel dimensions of 4.418 $×$ 4.418 $×$ 4.418 mm3. This served as input for Monte Carlo modelling and will be referred to as ‘geometry input’.

#### 2.2.2 Description of activity distribution

SPECT/CT images were acquired at 1 h, 4 h, 24 h, 48 h and 96 h post injection. Voxel-based heterogeneous activity distribution maps were derived after reconstructing the acquired SPECT/CT images and converting the counts in the reconstructed images to activity with the associated calibration factor. These activity maps consist of matrices of 128 $×$ 128 $×$ 90 with voxel dimensions of 4.418 $×$ 4.418 $×$ 4.418 mm3. The activity present in the whole field of view corresponding to the above time points was 2.08 GBq, 1.76 GBq, 1.27 GBq, 1 GBq and 0.65 GBq, respectively (Fig. 2). This served as ‘activity input’ for the corresponding geometry.

### 2.3 Reference dosimetry generation

The reference dosimetry or ‘ground truth’ was obtained by using Monte Carlo modelling with the Geant4-based code GATE. Voxel-based absorbed dose rate map determination requires the distribution of density as well as the activity distribution in the patient, at each time-point:
• The geometry input of each time point was used for that purpose. For each voxel, the Hounsfield units (HU) were converted to the material and density by the use of Schneider curve [
• Schneider W.
• Bortfeld T.
• Schlegel W.
Correlation between CT numbers and tissue parameters needed for Monte Carlo simulations of clinical dose distributions.
] available as HounsfieldMaterialGenerator in GATE [

GATE documentation (Voxelized source and phantom). https://opengate.readthedocs.io/en/latest/voxelized_source_and_phantom.html (accessed September 28, 2021).

].
• In addition, voxelized activity maps i.e. activity input for each time point derived from clinical acquisitions was used in GATE as source definition. Since all emissions contribute to the absorbed dose rate calculation, the radioactive decay along with the atomic de-excitation was simulated by using a 177Lu ion source defined by its atomic number (Z = 71), atomic weight (A = 177), ionic charge (Q = 0) and the excitation energy in keV (E = 0).
The DoseActor was used in GATE to calculate and store the absorbed dose rates in a given volume in a 3D matrix [

Sarrut D, Bardiès M, Boussion N, Freud N, Jan S, Létang J-M, et al. A review of the use and potential of the GATE Monte Carlo simulation code for radiation therapy and dosimetry applications. Med Phys 2014;41:064301.

] and was ‘attached to’ 3D patient volume. Since all particles were simulated, the number of simulated primaries was equal to the total activity present at each time point.
The generated dataset defines the ‘reference dosimetry’.

### 2.4 Monte Carlo based image generation

The participating clinical centre sends the host centre an image dataset (SPECT projections, corresponding CT, and reconstructed SPECT/CT) of a simple phantom (for example a uniform Jaszczak phantom, or an IEC phantom with or without spheres) acquired on their gamma camera. Along with this dataset, information regarding the gamma camera characteristics (such as crystal thickness, type of collimator, energy and spatial resolution), the activity settings (amount of activity in each compartment, date and time of activity calibration) and the acquisition protocol (acquisition time and duration, energy window information, type of orbit used, etc.) are required.
Following this, patient images are generated for each centre in accordance with their local acquisition protocols. The workflow adopted for image generation and data circulation is illustrated in Fig. 3.
Centre-specific SPECT/CT images are generated using GATE. The x and y axis of each projection corresponds to the sagittal (left to right) and axial (top to bottom) axis respectively. The simulated raw SPECT projections are incorporated into the template of the clinical images, and necessary tags in the DICOM (Digital Imaging and Communications in Medicine) header (for example: Largest Image Pixel Value, Smallest Image Pixel Value, CountsAccumulated) are adapted to the simulated dataset (using pydicom library [

Mason D. SU-E-T-33: Pydicom: An Open Source DICOM Library. Med Phys 2021;38(6Part10):3493-93.

] in a python3 script) so that they can be read by the clinical workstations present in the clinical centre.
Calibration images are modelled first. This has three potential advantages:
• It allows for the validation of the modelled gamma camera with a simple phantom,
• It validates the integration of a test image in the clinical centre image workstation,
• It enables the generation of a calibration factor corresponding to the simulated data set. This is important since the virtual gamma camera sensitivity is not necessarily strictly equal to that of a real gamma camera.

#### 2.4.1 Calibration image modelling

• i.
Simulation of calibration SPECT projections in GATE
A model based on the NEMA IEC calibration phantom, for which acquisitions were performed by the clinical centre, was developed for the simulations. Activity concentration for spheres (only the three largest ones) and the background along with the acquisition parameters were also obtained from the clinical acquisitions. The GE Infinia gamma camera model was derived from the work of Garcia et al. [
• Garcia M.-P.
• Villoing D.
• McKay E.
• Ferrer L.
• Cremonesi M.
• Botta F.
• et al.
TestDose: A nuclear medicine software based on Monte Carlo modeling for generating gamma camera acquisitions and dosimetry.
], and was adapted to the camera used for clinical acquisitions (⅜” NaI crystal size and medium energy collimator). The radial position, start angle and the angular step were extracted from the DICOM headers of clinical images, and used to simulate the gamma camera auto-contouring motion [
• Kayal G.
• Chauvin M.
• Mora-Ramirez E.
• Clayton N.
• Tran-Gia J.
• Lassmann M.
• et al.
]. The water-filled NEMA IEC phantom with the spheres of inner diameter 22 mm, 28 mm and 37 mm was modelled. For image modelling, only gamma emissions need to be considered. Therefore, contrary to the absorbed dose modelling, a weighted sum of gamma emissions (yield) was considered: 71.6 keV (0.15%), 112.9 keV (6.4%), 136.7 keV (0.05%), 208.4 keV (11%), 249.7 keV (0.21%) and 321.3 keV (0.22%) [

MIRD. https://www.nndc.bnl.gov/nudat2/mird/; 2020 (accessed September 30, 2021).

]. A total activity of 348.6 MBq was simulated and this corresponds to 5.66 $×$ 1010 primaries obtained as a product of activity in Bq, time per projection, number of projections and the total percentage of gamma emissions.
• ii.
Validation of simulated calibration projections
Simulated calibration projections were validated against experimental projections using two metrics:
• Flattened profiles (2D image to 1D profile):
1D profiles were drawn on the x-axis for each projection or 2D image by summing all the counts in the y-axis using a python script. This difference in magnitude was computed to assess the goodness of the simulated images.
• Gamma index (3D images)
The Gamma index, a very common metric used in the external beam radiotherapy [
• Low D.A.
• Harms W.B.
• Mutic S.
• Purdy J.A.
A technique for the quantitative evaluation of dose distributions.
] for absorbed dose distributions comparison, was employed to assess the similarity between experiments and simulations. This technique requires the user to specify two criteria for the evaluation: one is the absorbed dose difference (DD) i.e. is the acceptable difference in 3D absorbed dose distribution maps, and the other is the distance to agreement (DTA) i.e. the acceptable spatial difference between the compared images. From these, the gamma index passing rate (GIPR) and the average gamma was computed i.e. percentage and average number of points lying within the given DD/DTA acceptance criteria respectively.
The same metrics were used for the comparison of two data sets throughout the article (for SPECT projections, reconstructed SPECT/CT images, activity-indexed maps).
From the calibration projections, a ratio between the total counts in the measured and simulated projections was derived and utilised to adjust the patient projections. This is referred to as ‘normalisation’ in the Results section.
• iii.
Reconstruction of calibration projections
For the sake of the validation, the simulated SPECT projections were reconstructed in the same manner as the clinical dataset.
A calibration factor for the simulated data set was derived from the reconstructed SPECT/CT images using the following equation:
$CalibrationfactorCF=CountsVOIActivityAo×timeacq$
(1)

where CountsVOI, Ao and timeacq specifies the total counts in the volume of interest (VOI), the initial activity and the total time of acquisition respectively. Here, the whole field of view was considered as the VOI.

#### 2.4.2 Patient image modelling

• i.
Simulation of patient SPECT projections in GATE
The gamma camera model used for generating calibration images was used for patient image simulations. The patient model was segmented for soft tissues (liver, kidneys, spleen, tumour), bones, lungs and the remainder of the field of view. The segmentation was stored as a.nrrd (nearly raw raster data) file in Slicer, which is often used for scientific visualisation and image processing incorporating N-dimensional data. Tessellated mesh structures were generated for each volume of interest. This helped prevent patient and detector volume collision in the virtual GATE environment [
• Kayal G.
• Chauvin M.
• Mora-Ramirez E.
• Clayton N.
• Tran-Gia J.
• Lassmann M.
• et al.
]. The materials and the densities associated with each compartment or volume of interest were assigned in GATE. Heterogeneous activity distribution maps were generated for each time point. The 177Lu spectrum presented above for calibration image modelling was used. The number of primaries simulated from the first to the last time point ranged from 3.18 $×$ 1011 to 1.0 $×$ 1011 (Bq. s). The generation of patient SPECT images is described in more detail in Kayal et al. [
• Kayal G.
• Chauvin M.
• Vergara-Gil A.
• Clayton N.
• Ferrer L.
• Moalosi T.
• et al.
Generation of clinical 177Lu SPECT/CT images based on Monte Carlo simulation with GATE.
].
• ii.
Validation of simulated patient projections
The goodness of simulations was assessed using the same metrics previously mentioned for the calibration images – flattened profiles and gamma index (by specifying 2% DD and 1-pixel DTA criteria).
• iii.
Reconstruction of patient projections
Simulated SPECT projections were integrated into the template of the clinical images, and DICOM tags modified according to the procedure adopted for calibration images. The simulated patient SPECT/CT images were reconstructed using the same software and the same reconstruction parameters as the clinical dataset. These reconstructed SPECT/CT images have a dimension of 128 $×$ 128 $×$ 90 and voxel size of 4.418 $×$ 4.418 $×$ 4.418 mm3.

### 2.5 Validation procedure

Fig. 4 describes the workflow for evaluating the viability of the DosiTest project by comparing the absorbed dose rates generated directly from the input (referred to as ‘a’ in Fig. 4) to those obtained after reconstructing simulated SPECT/CT images and subsequent image-based clinical dosimetry (referred to as ‘d’ in Fig. 4).

#### 2.5.1 Activity-indexed images

• i.
Generation of simulated activity-indexed images.
The simulated patient reconstructed SPECT/CT images were fed into the OpenDose3D Slicer toolkit along with the derived simulation calibration factor, activity injected (6848 MBq) and date/time of activity administration. Activity-indexed maps from the simulated data set for each time point were generated.
• ii.
Activity comparison (Input vs generated activity images).
Activity input was compared to the simulated activity-indexed images obtained after the reconstruction and calibration of the simulated dataset at each time point using flattened profiles and gamma index. This is referred to as $'ActivityComparison'$ in Fig. 4.

#### 2.5.2 Absorbed Dose Rate (ADR)

• i.
From the simulated activity-indexed maps and the resampled CT, 3D absorbed dose rate maps for each time point were generated using GATE. This employed the $'GenerateGate'$ functionality in OpenDose3D Slicer module which creates the macros and data input files needed in GATE directly.
The GATE-generated absorbed dose rates (ADR) were then imported into OpenDose3D (through the built-in functionality $'ImportGate'$), which is capable of reading and interpreting the GATE-generated ADR for further dosimetric analysis.
• ii.
ADR comparison (Ground truth vs image-based).
As shown in Fig. 4, the ADR obtained directly from the input data (referred to as $'groundtruth'$ ADR) were compared to the ADR obtained from the simulated activity-indexed images (referred to as $'imagebased'$ ADR) via flattened profiles and gamma index.

#### 2.5.3 Integration of absorbed dose rates in time

Absorbed dose rate maps at each time point (both reference and image-based ADRs), were imported in OpenDose3D (using $'ImportGate'$). The segmentation performed on the clinical dataset (saved as.nrrd file) was imported for simulated datasets.
From the start of treatment (i.e. 0 h) to the first time point (i.e. 1 h), activity was assumed to be constant; from the first time point to infinity, a mono-exponential fitting algorithm was used, both for reference and image based datasets.

Absorbed doses for each volume of interest were obtained following the integration of absorbed dose rates. The relative difference between the absorbed doses obtained from reference (ADref) and image based (ADimg) ADRs was computed using the following equation:
$RelativeDifference%=ADref-ADimgADref$
(2)

## 3. Results

### 3.1 Reference dosimetry generation

The absorbed dose rate maps at all the time points were generated with GATE. The absorbed dose map at one hour post activity administration is presented in Fig. 5 along with the geometry input and activity input. Higher activity and therefore higher absorbed dose rates are seen for the tumour present in the liver. Reference absorbed dose rate maps obtained for all time points are presented in Fig. 6.

### 3.2 Monte Carlo based image generation

#### 3.2.1 Generation of calibration images

• i.
Simulated SPECT calibration projections.
Simulated calibration images were generated for the IEC phantom with three hot spheres and uniform background activity as can be seen in Fig. 7.
• ii.
Validation of simulated calibration projections.
Simulated calibration images were compared with the experimental calibration projections by drawing flattened profiles (Fig. 8). As can be seen, the largest sphere (37 mm diameter) has the highest activity and a peak corresponding to this sphere can also be seen in the flattened profiles at 0° gamma camera position (Fig. 8a). The GIPR and average gamma ($γavg$) between the clinical and simulated projections were 96.04% and 0.35 respectively.
• iii.
Reconstruction & generation of calibration factor.
The simulated calibration SPECT projections were incorporated in the DICOM header of clinical images and were reconstructed in Hermes clinical software. A simulated calibration factor of 6.46 counts per second (cps)/MBq was obtained using the Eq. (1). This value was further used to convert counts to activity in patient simulated dataset.
The calibration factor obtained from clinical images was 5.33 ± 0.27 cps/MBq. The simulated gamma camera exhibited higher gamma camera sensitivity, as is often observed. This highlights the importance of using a simulation based calibration factor for accurate activity quantification in patient simulated images.

#### 3.2.2 Generation of patient images

• i.
Simulated SPECT patient projections.
Simulated projections were generated in GATE using auto-contouring detector motion and tessellated mesh phantom for 1 h, 4 h, 24 h, 48 h and 96 h post activity administration. Fig. 9 displays these projections at different times post activity administration.
• ii.
Validation of simulated patient images.
The simulated SPECT projections were validated against experimental projections using flattened profiles and gamma index.
Flattened profiles between experiments and simulations were drawn after normalisation (Fig. 10), to take the difference of sensitivity between simulated and experimental images into account, and analysed for each projection angle from 0° to 360° with an angular step of 3° (therefore 60 projections/head). Gamma index was computed with 2% DD and 1-pixel DTA criteria. The GIPR and $γavg$ ranged from 95% to 98% and 0.28 to 0.35 respectively, over the different time points and projection angles between clinical and simulated projections, thereby illustrating the goodness of simulated patient SPECT projections.

### 3.3 Comparison of activities, absorbed dose rates and absorbed doses

#### 3.3.1 Activity-indexed images

• i.
Generation of simulated activity-indexed images.
A simulated calibration factor of 6.46 cps/MBq was used to convert from counts in reconstructed simulated SPECT projections to 3D activity-indexed maps. These simulated activity-indexed maps are presented in Fig. 11.
• ii.
Activity comparison (Input vs generated activity images).
The activity input was compared to the simulated activity-indexed images. The 1D summed profiles used for comparison are shown for the central slice in Fig. 12 as the tumour is clearly visible in this slice. However, profiles were drawn for all slices.
The highest peak in this figure refers to the activity in the tumour and the smaller peak corresponds to the spleen (Fig. 12). Due to lower activities in the 96 h post a.a., the activity maps are quite noisy. This impact can also be seen in the 96 h post a.a. profiles in Fig. 10 where SPECT projections are displayed.
Comparison of activity input and activity-indexed maps using average relative difference and the gamma index (GIPR and $γavg$) is shown in Table 1 for each time point. The relative differences are typically below 10%, except for the last time point activity images, which exhibit noise owing to the low activity. On the other hand, the gamma index reveals a significant degree of similarity between the two compared activity distribution maps, as seen in Table 1.
Table 1Comparison of the activity input and activity indexed maps using relative difference and gamma index metrics.
Time post a.a.Relative Diff (%)GIPR (%)$γavg$
1 h5.4697.310.28
4 h6.3497.710.26
24 h9.698.450.22
48 h7.9598.910.19
96 h10.1698.970.19

#### 3.3.2 Absorbed dose rate (ADR)

• i.
The image-based absorbed dose rate maps obtained from simulated activity-indexed maps are shown in Fig. 13. These are compared with the reference (or ‘ground truth’) ADRs in the following section.
• ii.
ADR comparison (Ground truth vs image based)
Ground truth or reference ADRs (Fig. 6) were compared against image-based ADRs (Fig. 13) to evaluate the similarities among them using flattened profiles (Fig. 14) and gamma index.
Table 2 shows the comparison based on gamma index and the computation of relative difference of the average absorbed dose rates for each time point. The presence of high GIPR values (and correspondingly low average gamma) demonstrated a significant degree of agreement between the absorbed dose rate maps. The relative difference in ADR across time points is less than 10%, therefore signifying minor statistical discrepancy.
Table 2Gamma metric comparison of the reference and image-based absorbed dose rate maps for each time point.
Time post a.a.Relative Diff (%)GIPR (%)$γavg$
1 h4.2795.630.3
4 h7.8896.640.3
24 h8.0397.420.26
48 h8.4698.200.22
96 h7.5998.040.23

The absorbed dose rates, both the reference and image-based were integrated after segmentation of volumes of interest and absorbed doses were computed. Table 3 shows the differences in the absorbed doses for the different volumes of interest namely liver, the kidneys, the spleen and the tumour present inside the liver.
Table 3Comparison of the absorbed doses (in Gy) obtained from reference (ADref) and image-based (ADimg) ADRs. The relative differences (Relative Diff.) between the AD are computed (using Eq. 2) for the volumes of interest.
Liver1.922.1713.16%
Left Kidney4.194.082.58%
Right Kidney3.493.480.32%
Spleen7.386.965.64%
Tumour16.2417.9010.22%

## 4. Discussion

This work presents for the first time the full description of DosiTest, a clinical dosimetry test based on Monte Carlo modelling designed to appraise the variability associated with the different steps of the clinical dosimetry workflow.
A model of a patient, with voxel-based geometry and activity, was created at various time-points after radiopharmaceutical administration. This allowed the generation of reference 3D absorbed dose rate maps for each time point from direct Monte Carlo modelling.
DosiTest's second step involves the generation of calibration and patient SPECT/CT datasets specific to each participating clinical centre.
Generation of calibration images from simple phantoms enabled the testing and validation of gamma camera models and specific acquisition procedures in a clinical centre. It also enabled the derivation of a simulated calibration factor.
The decision to use a ‘real’ patient dataset as a reference over ICRP adult phantoms [
• Menzel H.-G.
• Clement C.
• DeLuca P.
ICRP Publication 110. Realistic reference phantoms: an ICRP/ICRU joint effort. A report of adult reference computational phantoms.
] or NURBS-based XCAT phantom [
• Segars W.P.
• Sturgeon G.
• Mendonca S.
• Grimes J.
• Tsui B.M.W.
4D XCAT phantom for multimodality imaging research.
] was made for various reasons. The advantage of considering images obtained from a real patient is the availability of CT images at different time points. It allows the consideration of the image registration step (i.e. accounting for changes in patient anatomy and patient motion between imaging sessions), a non-trivial task that is likely to induce variability in the results. This is not possible when the same patient model (same geometry) is considered for all time points. Furthermore, it enables a realistic heterogeneous activity distribution within the patient model to be considered. The consequence, however, is that a full definition of the volumes of interest (reference segmentation) cannot be obtained.
The feasibility to generate these images (both patient and calibration) via Monte Carlo modelling with GATE was demonstrated.
One of the challenges of DosiTest is that in order to replicate the clinical situation, each clinical centre should perform dosimetry using their own tools and clinical image workstations. This means implicitly that simulated images have to be accepted by commercial image workstations as if they were “real” clinical images. The patient model, being generated from clinical images, already has a DICOM header. Therefore, it is possible to “copy/paste” raw projections in the clinical envelope, such that it can be processed by the clinical workstation.
Activity-indexed maps obtained after the reconstruction of simulated images and application of simulated calibration factor were in good agreement with the activity input (obtained with the respective calibration factor) for each time point. This highlights the need for the generation and use of calibration factors corresponding to a given dataset.
The 3D absorbed dose rates maps were further computed from the simulated image datasets using Monte Carlo modelling (referred to as the image-based ADR) and were compared with the reference absorbed dose rate maps obtained directly without imaging (referred to as the ground-truth ADR). These were found to be in good agreement as demonstrated in the Results section.
To derive absorbed doses for the volumes of interest from the absorbed dose rates (image-based and ground truth ADR), a user-dependent segmentation must be performed. This is certain to induce changes in the resulting absorbed doses. In our feasibility study, the same segmentation was used for both image-based and reference absorbed doses. Acceptable differences were found, even though the reason for some higher than expected absorbed doses should be further investigated (for example the liver where tumour is present).
We consider that the feasibility of DosiTest has been demonstrated. Further efforts will focus on the definition of checkpoints, i.e. intermediary steps where results can be exported and compared with a reference. This may involve extracting activity at various time points, evaluating the segmentation by providing the organs or tissues masses (even though there is no reference segmentation). It is clear that providing average absorbed doses (i.e. the end of the workflow) is not sufficient. Mora-Ramirez et al. [
• Mora‐Ramirez E.
• Santoro L.
• Cassol E.
• Ocampo‐Ramos J.C.
• Clayton N.
• Kayal G.
• et al.
Comparison of commercial dosimetric software platforms in patients treated with 177Lu-DOTATATE for peptide receptor radionuclide therapy.
] compared some of the clinical softwares available for dosimetry and concluded that each software implements its own dosimetric workflow, thereby complicating the comparison between workstations. Depending on the dosimetry procedure implemented in each clinical centre, it may be the case that only end-products (i.e. absorbed doses) are available, thereby limiting the strength of the analysis. However, given the current situation, where comparisons are mostly made on phantoms, and only question specific parts of the clinical dosimetry workflow (mostly that related to quantitative imaging), DosiTest is likely to bring extra information that is not currently available.

## 5. Conclusion

A proof of concept of the DosiTest project has been established. A patient model was defined, that includes a realistic definition of patient geometry and voxel-based activity, at different time points. The reference dosimetry (i.e. absorbed dose rates at the different times considered) was obtained from Monte Carlo modelling. Several image datasets were generated that allowed the integration into clinical image workstations and clinical dosimetry to be performed in each clinical centre according to their own local procedure.
Clinical dosimetry procedures are currently being tested within the IAEA CRP E23005, where results from at least eight clinical centres using the same clinical dosimetry software (Planet® Dose from DOSIsoft) are being compared.
The next step is the deployment of DosiTest in a limited number of clinical centres to start the appraisal of the various sources of variability in the clinical dosimetry procedure. This will enable the testing of DosiTest in more general situations where each clinical centre may use their own clinical dosimetry software and procedures.

### Acknowledgements

The patient images used were obtained as a part of IAEA Coordinated Research Project (CRP) on “Dosimetry in Radiopharmaceutical therapy for personalized patient treatment” (E2.30.05). The authors express sincere gratitude to Mr. Eli Stern from GE Healthcare, Israel for generating attenuation maps, Dr Peter Knoll and Dr Gian Luca Poli from IAEA, Austria for their active participation and constant support. This work was granted access to the HPC resources of CALMIP supercomputing centre under the allocation 2016-P19001. All the simulations have been performed with CALMIP and we thank them for their constant support.

### Funding

This work has been partially funded by the ENEN+ project that has received funding from the Euratom research and training Work programme 2016-2017-1 #75576.
OpenDose3D module in the Slicer3D toolkit has been developed as a part of the Medirad project, which received funding from the Euratom research and training programme 2014-2018 under grant agreement No. 755523.

## References

• Flux G.D.
• Sjogreen Gleisner K.
• Chiesa C.
• Lassmann M.
• Chouin N.
• Gear J.
• et al.
From fixed activities to personalized treatments in radionuclide therapy: lost in translation?.
Eur J Nucl Med Mol Imaging. 2018; 45: 152-154
• Sjögreen Gleisner K.
• Spezi E.
• Solny P.
• Gabina P.M.
• Cicone F.
• Stokke C.
• et al.
Variations in the practice of molecular radiotherapy and implementation of dosimetry: results from a European survey.
EJNMMI Phys. 2017; 4
• D’Arienzo M.
• Cazzato M.
• Cozzella M.L.
• Cox M.
• D’Andrea M.
• Fazio A.
• et al.
Gamma camera calibration and validation for quantitative SPECT imaging with 177Lu.
Appl Radiat Isot. 2016; 112: 156-164
• Gregory R.
• Flux G.
• Newbold K.
• Du Y.
• Moss L.
• et al.
SELIMETRY—a multicentre I-131 dosimetry trial: a clinical perspective.
BJR Suppl. 2017; 90: 20160637
• Wevrett J.
• Fenwick A.
• Scuffham J.
• Johansson L.
• Gear J.
• Schlögl S.
• et al.
Inter-comparison of quantitative imaging of lutetium-177 (177Lu) in European hospitals.
EJNMMI Phys. 2018; 5
• Zhao W.
• Esquinas P.L.
• Hou X.
• Uribe C.F.
• Gonzalez M.
• Beauregard J.-M.
• et al.
Determination of gamma camera calibration factors for quantitation of therapeutic radioisotopes.
EJNMMI Phys. 2018; 5
• Peters S.M.B.
• van der Werf N.R.
• Segbers M.
• van Velden F.H.P.
• Wierts R.
• Blokland K.A.K.
• et al.
Towards standardization of absolute SPECT/CT quantification: a multi-center and multi-vendor phantom study.
EJNMMI Phys. 2019; 6
• Taprogge J.
• Leek F.
• Schurrat T.
• Tran-Gia J.
• Vallot D.
• Bardiès M.
• et al.
Setting up a quantitative SPECT imaging network for a European multi-centre dosimetry study of radioiodine treatment for thyroid cancer as part of the MEDIRAD project.
EJNMMI Phys. 2020; 7
• Frezza A.
• Desport C.
• Uribe C.
• Zhao W.
• Celler A.
• Després P.
• et al.
Comprehensive SPECT/CT system characterization and calibration for 177Lu quantitative SPECT (QSPECT) with dead-time correction. EJNMMI.
Physics. 2020; 7https://doi.org/10.1186/s40658-020-0275-6
1. Dewaraja Y, Frey E, Sunderland J, Uribe C. Dosimetry Challenge https://therapy.snmmi.org/SNMMI-THERAPY/Dosimetry_Challenge.aspx (accessed October 13, 2021).

• Brolin G.
• Gleisner K.S.
• Ljungberg M.
Dynamic 99mTc-MAG3 renography: images for quality control obtained by combining pharmacokinetic modelling, an anthropomorphic computer phantom and Monte Carlo simulated scintillation camera imaging.
PMB. 2013; 58: 3145-3161
• Gustafsson J.
• Brolin G.
• Cox M.
• Ljungberg M.
• Johansson L.
• Gleisner K.S.
Uncertainty propagation for SPECT/CT-based renal dosimetry in 177Lu peptide receptor radionuclide therapy.
Phys Med Biol. 2015; 60: 8329-8346
• Brolin G.
• Gustafsson J.
• Ljungberg M.
• Gleisner K.S.
Pharmacokinetic digital phantoms for accuracy assessment of image-based dosimetry in (177)Lu-DOTATATE peptide receptor radionuclide therapy.
Phys Med Biol. 2015; 60: 6131-6149
2. Dositest project. http://www.dositest.org (accessed August 24, 2021).

3. Hapdey S, Soret M, Ferrer L, Koulibaly PM, Henriques J, Gardin I, et al. Quantification in SPECT: myth or reality? A multi-centric study. IEEE Symposium Conference Record Nuclear Science 2004, vol. 5, 2004, p. 3170–3 Vol. 5.

4. Ferrer L, McKay E, Lisbona A, Kraeber-Bodere F, Bardies M. DosiTest: Accuracy of a radio-immunotherapy dosimetry protocol. J Nucl Med 2009;50:1865–1865.

5. McKay E, Ferrer L, Barbet J, Bardies M. TestDose: Software for creating dosimetry problems. J Nucl Med 2009;50:1864–1864.

• Garcia M.-P.
• Villoing D.
• McKay E.
• Ferrer L.
• Cremonesi M.
• Botta F.
• et al.
TestDose: A nuclear medicine software based on Monte Carlo modeling for generating gamma camera acquisitions and dosimetry.
Med Phys. 2015; 42: 6885-6894
• Villoing D.
• Marcatili S.
• Garcia M.-P.
• Bardiès M.
Internal dosimetry with the Monte Carlo code GATE: validation using the ICRP/ICRU female reference computational model.
Phys Med Biol. 2017; 62: 1885-1904
• Costa G.C.A.
• Bonifácio D.A.B.
• Sarrut D.
• Cajgfinger T.
• Bardiès M.
Optimization of GATE simulations for whole-body planar scintigraphic acquisitions using the XCAT male phantom with 177Lu-DOTATATE biokinetics in a Siemens Symbia T2.
Phys Med. 2017; 42: 292-297
• Kayal G.
• Chauvin M.
• Vergara-Gil A.
• Clayton N.
• Ferrer L.
• Moalosi T.
• et al.
Generation of clinical 177Lu SPECT/CT images based on Monte Carlo simulation with GATE.
Phys Med. 2021; 85: 24-31
6. Dosimetry in molecular radiotherapy for personalized patient treatments. https://www.iaea.org/projects/crp/e23005 (accessed September 27, 2021).

7. Meijering EHW, Niessen WJ, Pluim JPW, Viergever MA. Quantitative Comparison of Sinc-Approximating Kernels for Medical Image Interpolation. Medical Image Computing and Computer-Assisted Intervention – MICCAI’99, Springer Berlin Heidelberg; 1999, p. 210–7.

• Vergara-Gil A.
• Amato E.
• Auditore L.
• Brenet M.
• Chauvin M.
• Clayton N.
• et al.
OpenDose3D: A free, collaborative 3D Slicer module for patient-specific dosimetry.
Eur J Nucl Med Mol Imaging. 2020; 47: S314-S315
• Pinter C.
• Lasso A.
• Wang A.n.
• Jaffray D.
• Fichtinger G.
SlicerRT: radiation therapy research toolkit for 3D Slicer.
Med Phys. 2012; 39: 6332-6338
• Schneider W.
• Bortfeld T.
• Schlegel W.
Correlation between CT numbers and tissue parameters needed for Monte Carlo simulations of clinical dose distributions.
Phys Med Biol. 2000; 45: 459-478
8. GATE documentation (Voxelized source and phantom). https://opengate.readthedocs.io/en/latest/voxelized_source_and_phantom.html (accessed September 28, 2021).

9. Sarrut D, Bardiès M, Boussion N, Freud N, Jan S, Létang J-M, et al. A review of the use and potential of the GATE Monte Carlo simulation code for radiation therapy and dosimetry applications. Med Phys 2014;41:064301.

• Kayal G.
• Chauvin M.
• Mora-Ramirez E.
• Clayton N.
• Tran-Gia J.
• Lassmann M.
• et al.
Modeling SPECT auto-contouring acquisition for 177Lu & 131I Molecular Radiotherapy using new developments in Geant4/GATE. 2022; 96: 101-113
10. MIRD. https://www.nndc.bnl.gov/nudat2/mird/; 2020 (accessed September 30, 2021).

• Low D.A.
• Harms W.B.
• Mutic S.
• Purdy J.A.
A technique for the quantitative evaluation of dose distributions.
Med Phys. 1998; 25: 656-661
11. Mason D. SU-E-T-33: Pydicom: An Open Source DICOM Library. Med Phys 2021;38(6Part10):3493-93.

• Dewaraja Y.K.
• Wilderman S.J.
• Ljungberg M.
• Koral K.F.
• Kaminiski M.S.
Accurate dosimetry in 131I radionuclide therapy using patient-specific, 3-dimensional methods for SPECT reconstruction and absorbed dose calculation.
J Nucl Med. 2005; 46: 840-849
• Menzel H.-G.
• Clement C.
• DeLuca P.
ICRP Publication 110. Realistic reference phantoms: an ICRP/ICRU joint effort. A report of adult reference computational phantoms.
Ann ICRP. 2009; 39: 1-164
• Segars W.P.
• Sturgeon G.
• Mendonca S.
• Grimes J.
• Tsui B.M.W.
4D XCAT phantom for multimodality imaging research.
Med Phys. 2010; 37: 4902-4915
• Mora‐Ramirez E.
• Santoro L.
• Cassol E.
• Ocampo‐Ramos J.C.
• Clayton N.
• Kayal G.
• et al.
Comparison of commercial dosimetric software platforms in patients treated with 177Lu-DOTATATE for peptide receptor radionuclide therapy.
Med Phys. 2020; 47: 4602-4615