- Email: [email protected]
An automated Monte Carlo QC system for volumetric modulated arc therapy: Possibilities and challenges
Physica Medica xxx (xxxx) xxx–xxx
Contents lists available at ScienceDirect
Physica Medica journal homepage: www.elsevier.com/locate/ejmp
Original paper
An automated Monte Carlo QC system for volumetric modulated arc therapy: Possibilities and challenges ⁎
R. Chakarovaa,b, , R. Cronholmc, M. Krantza, P. Anderssona, A. Hallqvistd a
Department of Medical Physics and Biomedical Engineering, Sahlgrenska University Hospital, Gothenburg, Sweden Department of Radiation Physics, Sahlgrenska Academy at the University of Gothenburg, Sweden c Department of Radiation Physics, Skåne University Hospital, Lund, Sweden d Department of Oncology, University of Gothenburg, Sweden b
A R T I C LE I N FO
A B S T R A C T
Keywords: Monte Carlo Patient specific VMAT QC Normalized dose difference
Purpose: To develop and implement an automated Monte Carlo (MC) system for patient specific VMAT quality control in a patient geometry that generates treatment planning system (TPS) compliant DICOM objects and includes a module for 3D analysis of dose deviations. Also, the aims were to recommend diagnose specific tolerance criteria and an evaluation procedure. Methods: The EGSnrc code package formed the basis for development of the MC system. The workflow consists of a number of modules connected to a TPS by means of manual DICOM exports and imports which were executed sequentially without user interaction. DVH comparison was performed in the TPS. In addition, MC- and TPS dose distributions were analysed by applying the normalized dose difference (NDD) formalism. NDD failure maps and a pass rate for a certain threshold were obtained. 170 clinical plans (prostate, thorax, head-and-neck and gynecological) were selected for analysis. Results: Agreement within 1.5% was found between clinical- and MC data for the mean dose to the target volumes and within 3% for parameters more sensitive to the shape of the DVH e.g. D98% PTV. Regarding the NDD analysis, tolerance criteria 2%/3 mm were established for prostate plans and 3%/3 mm for the rest of the cases. Conclusions: An automated MC system was developed and implemented. Evaluation procedure is recommended with NDD-analysis as a first step. For pass rate < 95%, the evaluation continues with comparison of DVH parameters. For deviations larger than 2%, a visual inspection of the clinical- and MC dose distributions is performed.
1. Introduction Technical advancements open possibilities for increasingly complex delivery techniques in external beam radiotherapy, such as volumetric modulated arc therapy (VMAT), which enables dynamic movements of the gantry and the multileaf collimator (MLC) and a variable dose rate during irradiation. Dose calculations in commercial treatment planning systems (TPS) are commonly based on approximations in terms of modelling the fluence and the dose deposition in media, which influence the calculation accuracy in heterogeneous media, especially for small apertures and steep dose gradients which are characteristic of VMAT. The non-reference dosimetry conditions require a thorough validation of the calculated dose distributions for each VMAT plan prior to patient treatment. The more advanced algorithm recently introduced, Acuros XB (AXB), is a deterministic algorithm explicitly
⁎
solving the linear Boltzmann transport equation (LBTE) [1]. The accuracy of the radiation transport is comparable to Monte Carlo (MC) stochastic methods indirectly obtaining the LBTE solution. However, the resulting AXB and MC dose distributions may differ since the description of the radiation source, the procedures of tissue segmentation and voxelization of patient geometry are not unified. Thus, MC can be used as a TPS independent method for dose validation. Development of the EGSnrc software [2] enables MC computation of dose distributions for a continuously variable beam configuration [3]. Extensive clinical record of VMAT plans verified by MC methods is reported in the literature [4]. The requirements are different for using the MC method as a quality control (QC) tool as opposed to being used as a research tool. In both cases, a careful validation of accelerator models, calibration and patient modelling is necessary. However, in the case of a QC tool, various steps in the calculation procedure cannot be
Corresponding author at: Department of Medical Physics and Biomedical Engineering, Sahlgrenska University Hospital, Gothenburg, Sweden. E-mail address: [email protected] (R. Chakarova).
https://doi.org/10.1016/j.ejmp.2018.03.010 Received 20 December 2017; Received in revised form 6 March 2018; Accepted 17 March 2018 1120-1797/ © 2018 Associazione Italiana di Fisica Medica. Published by Elsevier Ltd. All rights reserved.
Please cite this article as: Chakarova, R., Physica Medica (2018), https://doi.org/10.1016/j.ejmp.2018.03.010
Physica Medica xxx (xxxx) xxx–xxx
R. Chakarova et al.
DICOM RT files in the input directory. Treatment specific parameters were extracted from the RT Plan (RP) file, using pydicom [12], and written to instance specific copies of the template input files. The physical position of each MLC leaf was calculated from the light field position as described in the DICOM RP file by following the procedure in [13]. Based on CT images and RT structure (RS) file, a voxalized phantom was generated using the CTC-auto software; an interaction free python implementation of CTC-ask [14]. CTC-auto has the ability to handle any patient orientation including support structures if they are defined in the DICOM RS file. Generation of phantoms with voxel size comparable to CT resolution (1 mm) was possible but would increase the CPU time needed for a specific statistical dose accuracy. Resizing CT data may introduce variations in the material and density interpretation of the patient geometry affecting the dose distribution, especially in highly inhomogeneous regions like the nasal cavity as shown in [15]. Voxels ≤ 3 mm were recommended to reduce the voxel size effects [15]. Treatment site-specific resolution was implemented, following the dose calculation praxis in our hospital. Voxels of 2.5 × 2.5 × 3 mm3 size were used for gynecological and thorax treatments and of 2.5 × 2.5 × 2 mm3 size for prostate and head and neck (H&N). The dose grid information stored in the DICOM RT Dose (RD) files from the TPS was applied when preparing the phantom. Patient geometry was represented by nine tissues: air, lung, adipose and muscle tissues, as well as five bone tissues obtained by interpolation of bone mass density and composition between trabecular and cortical bone. The CT calibration curve was taken into account. Data for gold markers and titanium prosthesis were included in the segmentation table. BEAMnrc simulation for each input file of the instance was conducted starting from the phase space above the jaws and producing an intermediate phase space below the jaws in IAEA format to be further used in the DOSXYZnrc calculations. The initial phase spaces above the jaws contained a large enough number of particles, (about 8.108), to avoid particle recycling in the first simulation step. The number of particles in the intermediate phase space varied with jaw aperture, (e.g. 5.107–1.108 particles), to keep the particle density approximately constant. The MLCs were included by calling an instance of BEAMnrc as a shared library with an input file describing the planned dynamic movement. The fractional MU index of the intermediate phase space file was used to correlate to dynamic states of the same MU index. The MC system was capable of performing simulations by beam or by total plan depending on the input (DICOM RT DOSE by beam or total plan). Moreover, the MC system could handle any treatment technique described in a DICOM RP file. This was achieved by assuming that the last stated value of a parameter was also valid for control point where it was not explicitly stated. A small configurable number of histories was applied to run a DOSXYZnrc simulation for each input file in order to estimate the histories needed to reach a specified uncertainty. The final number of histories, (typically 5.107–9.107), was set so that the statistical uncertainty of the voxels occupying the isodose line corresponding to the prescription dose was 0.5% on average. The particle recycling rate of the intermediate phase space file was limited to three. The main DOSXYZnrc output was a 3ddose-file containing the resulting dose distribution. No statistical denoising was applied as post processing. Absolute dose in Gy was obtained by multiplying the MC dose by a calibration factor, the number of MUs and the instance specific backscatter correction factor (for Clinac iX only). Analytical description of the effect of the backscatter dose on the monitor chamber derived earlier was applied, assuming a linear dependence on the upper jaw opening [16]. Spencer-Attix restricted mass stopping-power ratios calculated earlier for the various tissues in the segmentation table were used to obtain the dose-to-water distributions [10]. It should be noted that a retrospective conversion of the MC data from dose-to-tissue to dose-to-water may increase the dose uncertainty, e.g. by introducing hot/cold spots
performed manually by MC knowledgeable users; evaluation criteria should be defined in detail and the results should assist decision-making about the reliability of the clinical dose distribution. MC system issues are successfully addressed in an MC based validation platform for Cyberknife and Tomotherapy treatments [5]. Although specific solutions are needed for the re-calculation of plans for various radiation techniques and treatment machines, the evaluation of the dose differences is a common task. The verification process usually includes visual comparison of dose-volume-histograms (DVH) for selected structures, comparison of physical parameters, (e.g. DVH constraints) and radiobiological indices [4–7]. A 3D analysis of calculated dose distributions, e.g. gamma evaluation, is not available in the TPS and the usage of a third party software is not straightforward [8,9]. The praxis of recalculation of VMAT plans on a homogeneous phantom is disadvantageous as the clinical importance of the results is not transparent. Independent dose calculations performed in the patient geometry as a part of the patient specific VMAT QC are more clinically relevant. The calculations can be supplemented by recommendations on how to perform dose analysis, on the selection of DVH parameters and on tolerance criteria to be applied. The aim of this work was to develop and implement an automated MC system for pre-treatment patient specific VMAT QC involving patient geometry that can generate TPS compliant DICOM objects and include a stand-alone module for 3D analysis of dose deviations. Also, the aims were to recommend diagnose specific tolerance criteria for DVH parameters and 3D dose differences in the irradiated volume. Identification of eventual limitations to a fully automated evaluation process was intended. 2. Materials and methods The EGSnrc code package v4-2.4.0 was used to develop the MC system [2,3]. The source code was modified as to allow more than 9 distinct media in the DOSXYZnrc simulation by specifying and reading the media matrix as space delimited. Workflow scripts were written exclusively in python and thus portable between different operating systems. Two layers of system configuration, defined in plain text files, were implemented to enhance the system flexibility. The global configuration regulated major parameters such as paths to input and output directories and the number of worker threads. The treatment machine specific configuration contained data about the MLC type, energy, fluence mode, calibration factors converting dose per incident particle to Gy/MU and dose to water conversion factors [10]. Machine specific template input files for BEAMnrc simulations and a global template input file for DOSXYZnrc calculations were provided, as well as phase space files above jaws. In this work, Varian accelerator models Clinac iX (Millenium 120 MLC), TrueBeam (Millenium 120 MLC) and TrueBeam (HDMLC), were used and validated for 6 MV flattened photon beams [11]. The BEAMnrc component module SYNCJAWS was used to model the collimator jaws, while SYNCMLC and SYNCHDMLC were used to model the Millenium 120 MLC and HDMLC respectively. The technical specifications from the vendor were followed when designing the MLC. Parameters like MLC bank offset, the position along the beam axis, the density for Millennium 120 MLC and the thickness for HDMLC, were determined by simulations and experimental comparison as described in [11]. Linear accelerators of different manufacturers and types can be handled simultaneously in the same workflow without requiring modifications of the framework. 2.1. MC workflow The workflow consisted of a number of modules (daemons) executed sequentially, automatable through crontabs, and connected to a TPS by means of manual DICOM exports and imports (Fig. 1). Each module, at the end of execution, generated data for the next module to process. A calculation instance was initiated by the appearance of 2
Physica Medica xxx (xxxx) xxx–xxx
R. Chakarova et al.
Fig. 1. The structure and workflow of the MC QC system.
calculation, such as tolerance levels and sub-region defined by dose limit were set using panel controls. The option to study NDD with respect to changes in dose grid resolution was enabled by adding an interpolation setting to the user interface. The visualization of the imported dose distributions, their dose difference, the NDD map as well as the map visualizing the failing voxels was enabled. A 2D distribution of the NDD values was shown as a diagram and the relative number of the voxels inside the tolerance criterion, i.e. pass-rate, was calculated (Fig. 2). Graphical and quantitative results were presented in separate windows.
and systematic errors [17,18]. Therefore, dose comparison in terms of dose-to-media is preferred if corresponding clinical calculations are available (e.g. AXB dose-to-medium mode). Dose-to-medium and dose-to-water MC dose distributions were written in DICOM format. DICOM objects were created by copying the TPS generated RP- and RD files and replacing the DICOM tags SOPInstaceUID and StudyInstanceUID. The MC dose matrix was stored in the RT Dose object. Metadata for position and shape of the matrix were updated if the data did not match the TPS generated metadata. In this way, when imported to the TPS, the MC generated DICOM files connected automatically to the primary image and structure set of that particular patient. The technical platform for the MC workflow was organized as an internal network of three high-performance (Haswell i7) computers installed as independent simulation servers along with the Linux based operating system Fedora 21. The typical CPU time to carry out a dose plan calculation on one computer with eight parallel threads was 3–5 h depending on the diagnosis.
2.3. Clinical material The functionality of the MC QC system was tested by performing MC calculations and subsequent analysis for treatment plans generated in Eclipse™ TPS. Single target region as well as treatment of anatomical regions with simultaneous multiple targets with different levels of the prescribed dose were included. Four cancer sites (170 patient plans in total) were considered; prostate (126 plans), gynecological (10 plans), H&N (13 plans) and thorax (21 plans). For these sites, a retrospective investigation was performed comparing the anisotropic analytical algorithm (AAA) and/ or AXB (v. 13.6.23) calculated- and MC obtained dose distributions. Treatment site-specific DVH estimates, such as DXX% (dose to XX% of the volume) or VZZ% (the volume receiving ZZ% of the prescribed dose), based on the recommended plan objectives and constraints for target volumes and OAR, were monitored for 170 plans. Representative sets of parameters were selected for further use in the 2D evaluation of the differences between MC and TPS dose distributions. The 3D NDD analysis was performed with the tolerance criteria 3%/3 mm and/or 2%/3 mm assuming a dose threshold of 20% of the prescribed dose.
2.2. Dose distribution comparison Two strategies were implemented; comparison within the TPS as well as evaluation in a separate analysis module developed in this work (Fig. 2). In the first case, visual inspection and comparisons of DVH parameters for target volumes and organs at risk (OAR) were performed. In addition, the dose distributions were imported into the stand-alone module for further 3D analysis (Fig. 1). The mathematical formalism of the normalized dose difference (NDD) method [19], an extension of the gamma evaluation concept, was implemented in the analysis module on MATLAB© (MathWorks, Boston, USA) based platform. Following the formalism, spatially varying normalization factors were calculated, i.e. the maximum allowed dose difference (MADD) at each point. The NDD value at a certain point of interest was obtained by the scaling of the dose difference at that point by the ratio of MADD to the predetermined dose acceptance tolerance. The NDD value has a sign, i.e. the information about which of the dose distributions is dominating, that was preserved. Voxels with absolute values smaller than the predetermined dose tolerance were assumed to pass both dose and spatial criteria. The module was used in an interactive mode. A user-friendly interface was developed. Parameters required to perform NDD
3. Implementation results and discussion DVH parameters for the clinical target volume (CTV) and the planning target volume (PTV), such as the mean doses, D98% PTV, D2% PTV as well as V90% for rectum as OAR, were selected for comparison in the cases of prostate cancer treatment. The average difference for the mean doses was close to zero (Fig. 3a). The DVH parameter D98%, obtained by AAA was systematically larger than the MC calculated, which may indicate a somewhat worse target coverage than planned. This parameter was more sensitive to the shape of the DVH than the mean 3
Physica Medica xxx (xxxx) xxx–xxx
R. Chakarova et al.
TPS
DVH: TPS vs. MC
MC
NDD analysis
Fig. 2. Evaluation strategies: Comparison in the TPS as well as evaluation in a stand-alone NDD analysis module.
applied. The average pass rate obtained was 96.2% with minimummaximum [88.3; 99.6]. Smaller deviations in the mean dose to the CTV and the PTV for higher pass rates have been observed (Fig. 4a). For example, the deviations were within 1.5% for pass rates larger than 95%. However, the opposite relation was not true, i.e. less than 1.5% deviation of the mean dose did not ensure high pass rate. The 3D NDD evaluation contained information complementary to the DVH analysis. Cases with pass rates below 95% were studied in detail. The typical
dose. The agreement between Acuros XB and MC results was within 0.5% for D98% PTV, D2% PTV (not shown in Fig. 3a). The average difference between AAA and MC was positive for V90% for rectum as well. However, the magnitude of a positive difference should not be of clinical concern for OAR. The NDD pass rates for 3%/3 mm tolerance criteria were found to be close to 100%. In order to enhance the differences between the TPSand MC dose distributions, a tolerance criteria of 2%/3 mm was
Fig. 3. Differences between TPS- and MC-calculated DVH parameters for target and OAR selected for treatment of (a) prostate cancer; (b) cancer in the thorax region; (c) H&N cancer and (d) gynecological cancer. Average differences for all patients in a particular group and the corresponding one standard deviation are shown.
4
Physica Medica xxx (xxxx) xxx–xxx
R. Chakarova et al.
Fig. 4. Deviations of the mean doses to clinical- and planned target volumes as a function of the pass rate estimated in the NDD analysis for treatment of (a) prostate cancer; (b) cancer in the thorax region; (c) cancer in the H&N region and (d) gynecological cancer.
Fig. 5. Comparison between MC- and TPS dose distributions for VMAT treatment of a prostate cancer. The dose range is set to [96–107%]. CTV has a volume of 79 cm3. Gold markers were present in CTV producing artefacts in the CT image. The shift in the TPS and MC calculated DVHs for CTV is illustrated.
correlation between the NDD and DVH results. For pass rates lower than 95%, the procedure should continue with DVH parameter comparison, and further with a visual inspection if deviations larger than 2% are detected. DVH parameters for the gross tumour volume (GTV) and the PTV as well as for lung and spinal cord (SC) as OAR were selected for comparison in the cases of treatment in thorax region (Fig. 3b). The mean differences were within 1% except for the minimum dose to GTV which is more sensitive to the choice of dose calculation algorithm. Average pass rate of 95.2 [85; 98.7] was obtained for tolerance criteria 3%/
causes of discrepancies between TPS- and MC dose distributions were differences in the tissue interpretation. For example, the presence of gold markers producing artefacts in the CT images introduced uncertainties in the tissue segmentation and consequently on the dose estimation (Fig. 5). Other sources of dose deviations were e.g. the presence of contrast in the bladder on the CT data or prosthesis artefacts overwritten in the TPS by help structures with assigned HU for water. In such cases, the advantage of the CTC-auto routine may be used to set the same material within exported help-structures as in the TPS. The evaluation procedure should start with NDD analysis due to the 5
Physica Medica xxx (xxxx) xxx–xxx
R. Chakarova et al.
routine procedure was recommended where the NDD method should be performed initially. For NDD pass rates < 95%, the evaluation should proceed with analysis of DVH parameters focusing on mean dose to target and OAR. If DVH parameter comparisons differs more than 2%; a visual inspection of dose distributions was recommended. A fully automated evaluation was hindered by artefacts in the CT images, presence of contrast in the bladder, dose to air included in the target volume, interpretation of HU in rectum etc. The MC framework is open source (available at: http://www.github. com/rickardcronholm/SAMC).
3 mm. The tendency of smaller deviations in the DVH parameters for higher pass rates was not as pronounced as for treatment of prostate cancer (Fig. 4b). The case with the exceptional low (85%) pass rate was further investigated by visual inspection. The treatment included two target volumes: a lateral one consisting mostly of lung tissue close to ribs and a medial one partly including trachea air. The MC results indicated lower target coverage when compared to AAA and AXB results, in particular, MC resulted in lower dose to the lung tissue included in the lateral PTV. The case manifested the potential for the MC method to evaluate the accuracy of TPS algorithms. For treatments in the H&N region, DVHs for the two target regions and parotid (P) and spinal cord as OAR were compared (Fig. 3c). Agreement within 2% was obtained for target mean doses as well as for the D98% parameters related to the shape of the DVH. The MC dose to parotid OAR was systematically lower compared to the TPS, which however was clinically acceptable. An average pass rate of 94.8 [92.3; 97.6] was obtained from the NDD analysis. Typical reasons for low pass rate were differences in the dose to the air cavities included in the target volumes and CT artefacts due to high density dental materials. Visual inspection of the target coverage showed that the AAA dose distribution usually dominated at the target boundary (the first and last target slices in cranial caudal direction). Also, a small number of voxels inside the targets were observed, where the AAA dose was larger. The maximum point dose for MC distribution was high, however the voxel volume was negligible. H&N regions are highly inhomogeneous and larger dose deviations were reported in the literature than for e.g. pelvis or abdomen [5] in line with our observation. Deviations for the mean doses to the two targets within 1.5% and a pass rate of 94.4 [83.4; 97.1] were obtained for treatment of gynecological cancer (Fig. 3d and Fig. 4d). Typical reasons for low pass rate were prosthesis, operational clips and interpretation of HU in the rectum. Dose deviations outside tolerances for the clinical plans analysed were explained by visual inspection of the dose distributions and patient anatomy. In an undefined discrepancy, the plan should be recalculated on a homogeneous phantom with subsequent measurements to exclude eventual deviations due to patient representation. Furthermore, some regions in patient geometry possibly contributing to the deviations could be assigned HU for water to evaluate the effect. If the problem persists, biological estimations such as tumour control probability, (TCP), and normal tissue complication probability, (NTCP), should be compared and a plan revision should be considered. Linking physical aspect of dose to clinical impact is not a trivial question. As pointed out in [5], the determination of warning levels and actions to be undertaken is challenging for a QC system implementation in the clinical praxis.
Acknowledgements Financial support of The Swedish Radiation Safety Authority and The Healthcare Committee, Region Västra Götaland are greatly acknowledged. Angela Lund is acknowledged for the Acuros XB data. References [1] Vassiliev ON, Wareing TA, McGhee J, Failla G, Salehpour MR, Mourtada F. Validation of a new grid-based Boltzmann equation solver for dose calculation in radiotherapy with photon beams. Phys Med Biol 2010;55:581–98. [2] Kawrakow I, Mainegra-Hing E, Rogers DWO, Tessier F, Walters BRB. The EGSnrc Code System: Monte Carlo Simulation of Electron and Photon Transport http://nrccnrc.github.io/EGSnrc/doc/pirs701-egsnrc.pdf. [3] Lobo J, Popescu IA. Two new DOSXYZnrc sources for 4D Monte Carlo simulations of continuously variable beam configurations, with applications to RapidArc, VMAT, TomoTherapy and CyberKnife. Phys Med Biol 2010;55:4431–43. [4] Popescu IA, Atwal P, Lobo J, Lucido J, McCurdy BMC. Patient-specific QA using 4D MC phase space predictions and EPID dosimetry. J Phys Conf Ser 2015;573(1):012004. [5] Wagner A, Crop F, Mirabel X, Tailly C, Reynaert N. Use of an in-house Monte Carlo platform to assess the clinical impact of algorithm-related dose differences on DVH constraints. Phys Med 2017;42:319–26. [6] Haga A, et al. Independent absorbed-dose calculation using the Monte Carlo algorithm in volumetric modulated arc therapy. Radiat Oncol 2014;9:1–9. [7] Andreou Maria, Karaiskos Pantelis, Kordolaimi Sofia, Koutsouveli Efi, Sandilos Panagiotis, Dimitriou Panagiotis, et al. Anatomy- vs. fluence-based planning for prostate cancer treatments using VMAT. Phys Med 2014;30(2):202–8. [8] Mukumoto N, et al. A preliminary study of in-house Monte Carlo simulations: an integrated Monte Carlo verification system. Int J Radiat Oncol Biol Phys 2009;75(2):571–9. [9] Agnew CE, McGarry CK. A tool to include gamma analysis software into a quality assurance program. Radiother Oncol 2016;118:568–73. [10] Siebers V, Keall PJ, Nahum AE, Mohan R. Converting absorbed dose to medium to absorbed dose to water for Monte Carlo based beam dose calculations. Phys Med Biol 2000;45:983–95. [11] Chakarova R, Cronholm R, Andersson P, Krantz M. Monte Carlo patient-specific pretreatment QA system for volumetric modulated arc therapy. Report 2017:13 ISSN: 2000–0456, available at www.stralsakerhetsmyndigheten.se/publikationer. [12] Mason D, Pydicom. https://github.com/darcymason/pydicom Access date: 201612-09. [13] Boyer AL. Li Shidong. Geometric analysis of light-field position of a multileaf collimator with curved ends. Med Phys 1997;24(5):757–62. [14] Ottosson RO, Behrens CF. CTC-ask: a new algorithm for conversion of CT numbers to tissue parameters for Monte Carlo dose calculations applying DICOM RS knowledge. Phys Med Biol 2011;56:N263–74. [15] Mora G, Pawlicki T, Maio A, Ma CM. Effect of Voxel size on Monte Carlo dose calculations for radiotherapy treatment planning. In: Kling A, editor. Advanced Monte Carlo for Radiation Physics, Particle transport simulation and applications. Berlin Heidelberg: Springer-Verlag; 2001. p. 549. [16] Hedin E, Bäck A, Chakarova R. Jaw position uncertainty and adjacent fields in breast cancer radiotherapy. JACMP 2015;16(6):240–51. [17] Ma CM, Li J. Dose specification for radiation therapy: dose to water or dose to medium? Phys Med Biol 2011;56(10):3073–89. [18] Andreo P. Dose to 'water-like' media or dose to tissue in MV photons radiotherapy treatment planning: still a matter of debate. Phys Med Biol 2015;60(1):309–37. [19] Jiang SB, Sharp GC, Neicu T, Berbeco RI, Flampouri S, Bortfeld T. On dose distribution comparison. Phys Med Biol 2006;51:759–76.
4. Conclusions An automated, flexible and portable MC system for independent QC of VMAT dose distributions, connected to a TPS by means of DICOM exports and imports, was developed and implemented for treatment plans for four cancer sites. Dose distribution comparison was performed in terms of DVH parameter comparison for target and OAR as well as in terms of 3D dose analysis by using the NDD method for selected dose and spatial tolerance criteria. Regarding the NDD analysis, tolerance criteria 2%/3 mm were suggested for prostate plans and 3%/3 mm for the rest of the cases. A
6
Contents lists available at ScienceDirect
Physica Medica journal homepage: www.elsevier.com/locate/ejmp
Original paper
An automated Monte Carlo QC system for volumetric modulated arc therapy: Possibilities and challenges ⁎
R. Chakarovaa,b, , R. Cronholmc, M. Krantza, P. Anderssona, A. Hallqvistd a
Department of Medical Physics and Biomedical Engineering, Sahlgrenska University Hospital, Gothenburg, Sweden Department of Radiation Physics, Sahlgrenska Academy at the University of Gothenburg, Sweden c Department of Radiation Physics, Skåne University Hospital, Lund, Sweden d Department of Oncology, University of Gothenburg, Sweden b
A R T I C LE I N FO
A B S T R A C T
Keywords: Monte Carlo Patient specific VMAT QC Normalized dose difference
Purpose: To develop and implement an automated Monte Carlo (MC) system for patient specific VMAT quality control in a patient geometry that generates treatment planning system (TPS) compliant DICOM objects and includes a module for 3D analysis of dose deviations. Also, the aims were to recommend diagnose specific tolerance criteria and an evaluation procedure. Methods: The EGSnrc code package formed the basis for development of the MC system. The workflow consists of a number of modules connected to a TPS by means of manual DICOM exports and imports which were executed sequentially without user interaction. DVH comparison was performed in the TPS. In addition, MC- and TPS dose distributions were analysed by applying the normalized dose difference (NDD) formalism. NDD failure maps and a pass rate for a certain threshold were obtained. 170 clinical plans (prostate, thorax, head-and-neck and gynecological) were selected for analysis. Results: Agreement within 1.5% was found between clinical- and MC data for the mean dose to the target volumes and within 3% for parameters more sensitive to the shape of the DVH e.g. D98% PTV. Regarding the NDD analysis, tolerance criteria 2%/3 mm were established for prostate plans and 3%/3 mm for the rest of the cases. Conclusions: An automated MC system was developed and implemented. Evaluation procedure is recommended with NDD-analysis as a first step. For pass rate < 95%, the evaluation continues with comparison of DVH parameters. For deviations larger than 2%, a visual inspection of the clinical- and MC dose distributions is performed.
1. Introduction Technical advancements open possibilities for increasingly complex delivery techniques in external beam radiotherapy, such as volumetric modulated arc therapy (VMAT), which enables dynamic movements of the gantry and the multileaf collimator (MLC) and a variable dose rate during irradiation. Dose calculations in commercial treatment planning systems (TPS) are commonly based on approximations in terms of modelling the fluence and the dose deposition in media, which influence the calculation accuracy in heterogeneous media, especially for small apertures and steep dose gradients which are characteristic of VMAT. The non-reference dosimetry conditions require a thorough validation of the calculated dose distributions for each VMAT plan prior to patient treatment. The more advanced algorithm recently introduced, Acuros XB (AXB), is a deterministic algorithm explicitly
⁎
solving the linear Boltzmann transport equation (LBTE) [1]. The accuracy of the radiation transport is comparable to Monte Carlo (MC) stochastic methods indirectly obtaining the LBTE solution. However, the resulting AXB and MC dose distributions may differ since the description of the radiation source, the procedures of tissue segmentation and voxelization of patient geometry are not unified. Thus, MC can be used as a TPS independent method for dose validation. Development of the EGSnrc software [2] enables MC computation of dose distributions for a continuously variable beam configuration [3]. Extensive clinical record of VMAT plans verified by MC methods is reported in the literature [4]. The requirements are different for using the MC method as a quality control (QC) tool as opposed to being used as a research tool. In both cases, a careful validation of accelerator models, calibration and patient modelling is necessary. However, in the case of a QC tool, various steps in the calculation procedure cannot be
Corresponding author at: Department of Medical Physics and Biomedical Engineering, Sahlgrenska University Hospital, Gothenburg, Sweden. E-mail address: [email protected] (R. Chakarova).
https://doi.org/10.1016/j.ejmp.2018.03.010 Received 20 December 2017; Received in revised form 6 March 2018; Accepted 17 March 2018 1120-1797/ © 2018 Associazione Italiana di Fisica Medica. Published by Elsevier Ltd. All rights reserved.
Please cite this article as: Chakarova, R., Physica Medica (2018), https://doi.org/10.1016/j.ejmp.2018.03.010
Physica Medica xxx (xxxx) xxx–xxx
R. Chakarova et al.
DICOM RT files in the input directory. Treatment specific parameters were extracted from the RT Plan (RP) file, using pydicom [12], and written to instance specific copies of the template input files. The physical position of each MLC leaf was calculated from the light field position as described in the DICOM RP file by following the procedure in [13]. Based on CT images and RT structure (RS) file, a voxalized phantom was generated using the CTC-auto software; an interaction free python implementation of CTC-ask [14]. CTC-auto has the ability to handle any patient orientation including support structures if they are defined in the DICOM RS file. Generation of phantoms with voxel size comparable to CT resolution (1 mm) was possible but would increase the CPU time needed for a specific statistical dose accuracy. Resizing CT data may introduce variations in the material and density interpretation of the patient geometry affecting the dose distribution, especially in highly inhomogeneous regions like the nasal cavity as shown in [15]. Voxels ≤ 3 mm were recommended to reduce the voxel size effects [15]. Treatment site-specific resolution was implemented, following the dose calculation praxis in our hospital. Voxels of 2.5 × 2.5 × 3 mm3 size were used for gynecological and thorax treatments and of 2.5 × 2.5 × 2 mm3 size for prostate and head and neck (H&N). The dose grid information stored in the DICOM RT Dose (RD) files from the TPS was applied when preparing the phantom. Patient geometry was represented by nine tissues: air, lung, adipose and muscle tissues, as well as five bone tissues obtained by interpolation of bone mass density and composition between trabecular and cortical bone. The CT calibration curve was taken into account. Data for gold markers and titanium prosthesis were included in the segmentation table. BEAMnrc simulation for each input file of the instance was conducted starting from the phase space above the jaws and producing an intermediate phase space below the jaws in IAEA format to be further used in the DOSXYZnrc calculations. The initial phase spaces above the jaws contained a large enough number of particles, (about 8.108), to avoid particle recycling in the first simulation step. The number of particles in the intermediate phase space varied with jaw aperture, (e.g. 5.107–1.108 particles), to keep the particle density approximately constant. The MLCs were included by calling an instance of BEAMnrc as a shared library with an input file describing the planned dynamic movement. The fractional MU index of the intermediate phase space file was used to correlate to dynamic states of the same MU index. The MC system was capable of performing simulations by beam or by total plan depending on the input (DICOM RT DOSE by beam or total plan). Moreover, the MC system could handle any treatment technique described in a DICOM RP file. This was achieved by assuming that the last stated value of a parameter was also valid for control point where it was not explicitly stated. A small configurable number of histories was applied to run a DOSXYZnrc simulation for each input file in order to estimate the histories needed to reach a specified uncertainty. The final number of histories, (typically 5.107–9.107), was set so that the statistical uncertainty of the voxels occupying the isodose line corresponding to the prescription dose was 0.5% on average. The particle recycling rate of the intermediate phase space file was limited to three. The main DOSXYZnrc output was a 3ddose-file containing the resulting dose distribution. No statistical denoising was applied as post processing. Absolute dose in Gy was obtained by multiplying the MC dose by a calibration factor, the number of MUs and the instance specific backscatter correction factor (for Clinac iX only). Analytical description of the effect of the backscatter dose on the monitor chamber derived earlier was applied, assuming a linear dependence on the upper jaw opening [16]. Spencer-Attix restricted mass stopping-power ratios calculated earlier for the various tissues in the segmentation table were used to obtain the dose-to-water distributions [10]. It should be noted that a retrospective conversion of the MC data from dose-to-tissue to dose-to-water may increase the dose uncertainty, e.g. by introducing hot/cold spots
performed manually by MC knowledgeable users; evaluation criteria should be defined in detail and the results should assist decision-making about the reliability of the clinical dose distribution. MC system issues are successfully addressed in an MC based validation platform for Cyberknife and Tomotherapy treatments [5]. Although specific solutions are needed for the re-calculation of plans for various radiation techniques and treatment machines, the evaluation of the dose differences is a common task. The verification process usually includes visual comparison of dose-volume-histograms (DVH) for selected structures, comparison of physical parameters, (e.g. DVH constraints) and radiobiological indices [4–7]. A 3D analysis of calculated dose distributions, e.g. gamma evaluation, is not available in the TPS and the usage of a third party software is not straightforward [8,9]. The praxis of recalculation of VMAT plans on a homogeneous phantom is disadvantageous as the clinical importance of the results is not transparent. Independent dose calculations performed in the patient geometry as a part of the patient specific VMAT QC are more clinically relevant. The calculations can be supplemented by recommendations on how to perform dose analysis, on the selection of DVH parameters and on tolerance criteria to be applied. The aim of this work was to develop and implement an automated MC system for pre-treatment patient specific VMAT QC involving patient geometry that can generate TPS compliant DICOM objects and include a stand-alone module for 3D analysis of dose deviations. Also, the aims were to recommend diagnose specific tolerance criteria for DVH parameters and 3D dose differences in the irradiated volume. Identification of eventual limitations to a fully automated evaluation process was intended. 2. Materials and methods The EGSnrc code package v4-2.4.0 was used to develop the MC system [2,3]. The source code was modified as to allow more than 9 distinct media in the DOSXYZnrc simulation by specifying and reading the media matrix as space delimited. Workflow scripts were written exclusively in python and thus portable between different operating systems. Two layers of system configuration, defined in plain text files, were implemented to enhance the system flexibility. The global configuration regulated major parameters such as paths to input and output directories and the number of worker threads. The treatment machine specific configuration contained data about the MLC type, energy, fluence mode, calibration factors converting dose per incident particle to Gy/MU and dose to water conversion factors [10]. Machine specific template input files for BEAMnrc simulations and a global template input file for DOSXYZnrc calculations were provided, as well as phase space files above jaws. In this work, Varian accelerator models Clinac iX (Millenium 120 MLC), TrueBeam (Millenium 120 MLC) and TrueBeam (HDMLC), were used and validated for 6 MV flattened photon beams [11]. The BEAMnrc component module SYNCJAWS was used to model the collimator jaws, while SYNCMLC and SYNCHDMLC were used to model the Millenium 120 MLC and HDMLC respectively. The technical specifications from the vendor were followed when designing the MLC. Parameters like MLC bank offset, the position along the beam axis, the density for Millennium 120 MLC and the thickness for HDMLC, were determined by simulations and experimental comparison as described in [11]. Linear accelerators of different manufacturers and types can be handled simultaneously in the same workflow without requiring modifications of the framework. 2.1. MC workflow The workflow consisted of a number of modules (daemons) executed sequentially, automatable through crontabs, and connected to a TPS by means of manual DICOM exports and imports (Fig. 1). Each module, at the end of execution, generated data for the next module to process. A calculation instance was initiated by the appearance of 2
Physica Medica xxx (xxxx) xxx–xxx
R. Chakarova et al.
Fig. 1. The structure and workflow of the MC QC system.
calculation, such as tolerance levels and sub-region defined by dose limit were set using panel controls. The option to study NDD with respect to changes in dose grid resolution was enabled by adding an interpolation setting to the user interface. The visualization of the imported dose distributions, their dose difference, the NDD map as well as the map visualizing the failing voxels was enabled. A 2D distribution of the NDD values was shown as a diagram and the relative number of the voxels inside the tolerance criterion, i.e. pass-rate, was calculated (Fig. 2). Graphical and quantitative results were presented in separate windows.
and systematic errors [17,18]. Therefore, dose comparison in terms of dose-to-media is preferred if corresponding clinical calculations are available (e.g. AXB dose-to-medium mode). Dose-to-medium and dose-to-water MC dose distributions were written in DICOM format. DICOM objects were created by copying the TPS generated RP- and RD files and replacing the DICOM tags SOPInstaceUID and StudyInstanceUID. The MC dose matrix was stored in the RT Dose object. Metadata for position and shape of the matrix were updated if the data did not match the TPS generated metadata. In this way, when imported to the TPS, the MC generated DICOM files connected automatically to the primary image and structure set of that particular patient. The technical platform for the MC workflow was organized as an internal network of three high-performance (Haswell i7) computers installed as independent simulation servers along with the Linux based operating system Fedora 21. The typical CPU time to carry out a dose plan calculation on one computer with eight parallel threads was 3–5 h depending on the diagnosis.
2.3. Clinical material The functionality of the MC QC system was tested by performing MC calculations and subsequent analysis for treatment plans generated in Eclipse™ TPS. Single target region as well as treatment of anatomical regions with simultaneous multiple targets with different levels of the prescribed dose were included. Four cancer sites (170 patient plans in total) were considered; prostate (126 plans), gynecological (10 plans), H&N (13 plans) and thorax (21 plans). For these sites, a retrospective investigation was performed comparing the anisotropic analytical algorithm (AAA) and/ or AXB (v. 13.6.23) calculated- and MC obtained dose distributions. Treatment site-specific DVH estimates, such as DXX% (dose to XX% of the volume) or VZZ% (the volume receiving ZZ% of the prescribed dose), based on the recommended plan objectives and constraints for target volumes and OAR, were monitored for 170 plans. Representative sets of parameters were selected for further use in the 2D evaluation of the differences between MC and TPS dose distributions. The 3D NDD analysis was performed with the tolerance criteria 3%/3 mm and/or 2%/3 mm assuming a dose threshold of 20% of the prescribed dose.
2.2. Dose distribution comparison Two strategies were implemented; comparison within the TPS as well as evaluation in a separate analysis module developed in this work (Fig. 2). In the first case, visual inspection and comparisons of DVH parameters for target volumes and organs at risk (OAR) were performed. In addition, the dose distributions were imported into the stand-alone module for further 3D analysis (Fig. 1). The mathematical formalism of the normalized dose difference (NDD) method [19], an extension of the gamma evaluation concept, was implemented in the analysis module on MATLAB© (MathWorks, Boston, USA) based platform. Following the formalism, spatially varying normalization factors were calculated, i.e. the maximum allowed dose difference (MADD) at each point. The NDD value at a certain point of interest was obtained by the scaling of the dose difference at that point by the ratio of MADD to the predetermined dose acceptance tolerance. The NDD value has a sign, i.e. the information about which of the dose distributions is dominating, that was preserved. Voxels with absolute values smaller than the predetermined dose tolerance were assumed to pass both dose and spatial criteria. The module was used in an interactive mode. A user-friendly interface was developed. Parameters required to perform NDD
3. Implementation results and discussion DVH parameters for the clinical target volume (CTV) and the planning target volume (PTV), such as the mean doses, D98% PTV, D2% PTV as well as V90% for rectum as OAR, were selected for comparison in the cases of prostate cancer treatment. The average difference for the mean doses was close to zero (Fig. 3a). The DVH parameter D98%, obtained by AAA was systematically larger than the MC calculated, which may indicate a somewhat worse target coverage than planned. This parameter was more sensitive to the shape of the DVH than the mean 3
Physica Medica xxx (xxxx) xxx–xxx
R. Chakarova et al.
TPS
DVH: TPS vs. MC
MC
NDD analysis
Fig. 2. Evaluation strategies: Comparison in the TPS as well as evaluation in a stand-alone NDD analysis module.
applied. The average pass rate obtained was 96.2% with minimummaximum [88.3; 99.6]. Smaller deviations in the mean dose to the CTV and the PTV for higher pass rates have been observed (Fig. 4a). For example, the deviations were within 1.5% for pass rates larger than 95%. However, the opposite relation was not true, i.e. less than 1.5% deviation of the mean dose did not ensure high pass rate. The 3D NDD evaluation contained information complementary to the DVH analysis. Cases with pass rates below 95% were studied in detail. The typical
dose. The agreement between Acuros XB and MC results was within 0.5% for D98% PTV, D2% PTV (not shown in Fig. 3a). The average difference between AAA and MC was positive for V90% for rectum as well. However, the magnitude of a positive difference should not be of clinical concern for OAR. The NDD pass rates for 3%/3 mm tolerance criteria were found to be close to 100%. In order to enhance the differences between the TPSand MC dose distributions, a tolerance criteria of 2%/3 mm was
Fig. 3. Differences between TPS- and MC-calculated DVH parameters for target and OAR selected for treatment of (a) prostate cancer; (b) cancer in the thorax region; (c) H&N cancer and (d) gynecological cancer. Average differences for all patients in a particular group and the corresponding one standard deviation are shown.
4
Physica Medica xxx (xxxx) xxx–xxx
R. Chakarova et al.
Fig. 4. Deviations of the mean doses to clinical- and planned target volumes as a function of the pass rate estimated in the NDD analysis for treatment of (a) prostate cancer; (b) cancer in the thorax region; (c) cancer in the H&N region and (d) gynecological cancer.
Fig. 5. Comparison between MC- and TPS dose distributions for VMAT treatment of a prostate cancer. The dose range is set to [96–107%]. CTV has a volume of 79 cm3. Gold markers were present in CTV producing artefacts in the CT image. The shift in the TPS and MC calculated DVHs for CTV is illustrated.
correlation between the NDD and DVH results. For pass rates lower than 95%, the procedure should continue with DVH parameter comparison, and further with a visual inspection if deviations larger than 2% are detected. DVH parameters for the gross tumour volume (GTV) and the PTV as well as for lung and spinal cord (SC) as OAR were selected for comparison in the cases of treatment in thorax region (Fig. 3b). The mean differences were within 1% except for the minimum dose to GTV which is more sensitive to the choice of dose calculation algorithm. Average pass rate of 95.2 [85; 98.7] was obtained for tolerance criteria 3%/
causes of discrepancies between TPS- and MC dose distributions were differences in the tissue interpretation. For example, the presence of gold markers producing artefacts in the CT images introduced uncertainties in the tissue segmentation and consequently on the dose estimation (Fig. 5). Other sources of dose deviations were e.g. the presence of contrast in the bladder on the CT data or prosthesis artefacts overwritten in the TPS by help structures with assigned HU for water. In such cases, the advantage of the CTC-auto routine may be used to set the same material within exported help-structures as in the TPS. The evaluation procedure should start with NDD analysis due to the 5
Physica Medica xxx (xxxx) xxx–xxx
R. Chakarova et al.
routine procedure was recommended where the NDD method should be performed initially. For NDD pass rates < 95%, the evaluation should proceed with analysis of DVH parameters focusing on mean dose to target and OAR. If DVH parameter comparisons differs more than 2%; a visual inspection of dose distributions was recommended. A fully automated evaluation was hindered by artefacts in the CT images, presence of contrast in the bladder, dose to air included in the target volume, interpretation of HU in rectum etc. The MC framework is open source (available at: http://www.github. com/rickardcronholm/SAMC).
3 mm. The tendency of smaller deviations in the DVH parameters for higher pass rates was not as pronounced as for treatment of prostate cancer (Fig. 4b). The case with the exceptional low (85%) pass rate was further investigated by visual inspection. The treatment included two target volumes: a lateral one consisting mostly of lung tissue close to ribs and a medial one partly including trachea air. The MC results indicated lower target coverage when compared to AAA and AXB results, in particular, MC resulted in lower dose to the lung tissue included in the lateral PTV. The case manifested the potential for the MC method to evaluate the accuracy of TPS algorithms. For treatments in the H&N region, DVHs for the two target regions and parotid (P) and spinal cord as OAR were compared (Fig. 3c). Agreement within 2% was obtained for target mean doses as well as for the D98% parameters related to the shape of the DVH. The MC dose to parotid OAR was systematically lower compared to the TPS, which however was clinically acceptable. An average pass rate of 94.8 [92.3; 97.6] was obtained from the NDD analysis. Typical reasons for low pass rate were differences in the dose to the air cavities included in the target volumes and CT artefacts due to high density dental materials. Visual inspection of the target coverage showed that the AAA dose distribution usually dominated at the target boundary (the first and last target slices in cranial caudal direction). Also, a small number of voxels inside the targets were observed, where the AAA dose was larger. The maximum point dose for MC distribution was high, however the voxel volume was negligible. H&N regions are highly inhomogeneous and larger dose deviations were reported in the literature than for e.g. pelvis or abdomen [5] in line with our observation. Deviations for the mean doses to the two targets within 1.5% and a pass rate of 94.4 [83.4; 97.1] were obtained for treatment of gynecological cancer (Fig. 3d and Fig. 4d). Typical reasons for low pass rate were prosthesis, operational clips and interpretation of HU in the rectum. Dose deviations outside tolerances for the clinical plans analysed were explained by visual inspection of the dose distributions and patient anatomy. In an undefined discrepancy, the plan should be recalculated on a homogeneous phantom with subsequent measurements to exclude eventual deviations due to patient representation. Furthermore, some regions in patient geometry possibly contributing to the deviations could be assigned HU for water to evaluate the effect. If the problem persists, biological estimations such as tumour control probability, (TCP), and normal tissue complication probability, (NTCP), should be compared and a plan revision should be considered. Linking physical aspect of dose to clinical impact is not a trivial question. As pointed out in [5], the determination of warning levels and actions to be undertaken is challenging for a QC system implementation in the clinical praxis.
Acknowledgements Financial support of The Swedish Radiation Safety Authority and The Healthcare Committee, Region Västra Götaland are greatly acknowledged. Angela Lund is acknowledged for the Acuros XB data. References [1] Vassiliev ON, Wareing TA, McGhee J, Failla G, Salehpour MR, Mourtada F. Validation of a new grid-based Boltzmann equation solver for dose calculation in radiotherapy with photon beams. Phys Med Biol 2010;55:581–98. [2] Kawrakow I, Mainegra-Hing E, Rogers DWO, Tessier F, Walters BRB. The EGSnrc Code System: Monte Carlo Simulation of Electron and Photon Transport http://nrccnrc.github.io/EGSnrc/doc/pirs701-egsnrc.pdf. [3] Lobo J, Popescu IA. Two new DOSXYZnrc sources for 4D Monte Carlo simulations of continuously variable beam configurations, with applications to RapidArc, VMAT, TomoTherapy and CyberKnife. Phys Med Biol 2010;55:4431–43. [4] Popescu IA, Atwal P, Lobo J, Lucido J, McCurdy BMC. Patient-specific QA using 4D MC phase space predictions and EPID dosimetry. J Phys Conf Ser 2015;573(1):012004. [5] Wagner A, Crop F, Mirabel X, Tailly C, Reynaert N. Use of an in-house Monte Carlo platform to assess the clinical impact of algorithm-related dose differences on DVH constraints. Phys Med 2017;42:319–26. [6] Haga A, et al. Independent absorbed-dose calculation using the Monte Carlo algorithm in volumetric modulated arc therapy. Radiat Oncol 2014;9:1–9. [7] Andreou Maria, Karaiskos Pantelis, Kordolaimi Sofia, Koutsouveli Efi, Sandilos Panagiotis, Dimitriou Panagiotis, et al. Anatomy- vs. fluence-based planning for prostate cancer treatments using VMAT. Phys Med 2014;30(2):202–8. [8] Mukumoto N, et al. A preliminary study of in-house Monte Carlo simulations: an integrated Monte Carlo verification system. Int J Radiat Oncol Biol Phys 2009;75(2):571–9. [9] Agnew CE, McGarry CK. A tool to include gamma analysis software into a quality assurance program. Radiother Oncol 2016;118:568–73. [10] Siebers V, Keall PJ, Nahum AE, Mohan R. Converting absorbed dose to medium to absorbed dose to water for Monte Carlo based beam dose calculations. Phys Med Biol 2000;45:983–95. [11] Chakarova R, Cronholm R, Andersson P, Krantz M. Monte Carlo patient-specific pretreatment QA system for volumetric modulated arc therapy. Report 2017:13 ISSN: 2000–0456, available at www.stralsakerhetsmyndigheten.se/publikationer. [12] Mason D, Pydicom. https://github.com/darcymason/pydicom Access date: 201612-09. [13] Boyer AL. Li Shidong. Geometric analysis of light-field position of a multileaf collimator with curved ends. Med Phys 1997;24(5):757–62. [14] Ottosson RO, Behrens CF. CTC-ask: a new algorithm for conversion of CT numbers to tissue parameters for Monte Carlo dose calculations applying DICOM RS knowledge. Phys Med Biol 2011;56:N263–74. [15] Mora G, Pawlicki T, Maio A, Ma CM. Effect of Voxel size on Monte Carlo dose calculations for radiotherapy treatment planning. In: Kling A, editor. Advanced Monte Carlo for Radiation Physics, Particle transport simulation and applications. Berlin Heidelberg: Springer-Verlag; 2001. p. 549. [16] Hedin E, Bäck A, Chakarova R. Jaw position uncertainty and adjacent fields in breast cancer radiotherapy. JACMP 2015;16(6):240–51. [17] Ma CM, Li J. Dose specification for radiation therapy: dose to water or dose to medium? Phys Med Biol 2011;56(10):3073–89. [18] Andreo P. Dose to 'water-like' media or dose to tissue in MV photons radiotherapy treatment planning: still a matter of debate. Phys Med Biol 2015;60(1):309–37. [19] Jiang SB, Sharp GC, Neicu T, Berbeco RI, Flampouri S, Bortfeld T. On dose distribution comparison. Phys Med Biol 2006;51:759–76.
4. Conclusions An automated, flexible and portable MC system for independent QC of VMAT dose distributions, connected to a TPS by means of DICOM exports and imports, was developed and implemented for treatment plans for four cancer sites. Dose distribution comparison was performed in terms of DVH parameter comparison for target and OAR as well as in terms of 3D dose analysis by using the NDD method for selected dose and spatial tolerance criteria. Regarding the NDD analysis, tolerance criteria 2%/3 mm were suggested for prostate plans and 3%/3 mm for the rest of the cases. A
6