Reducing variability among treatment machines using knowledge‐based planning for head and neck, pancreatic, and rectal cancer

Abstract Purpose This study aimed to assess dosimetric indices of RapidPlan model‐based plans for different energies (6, 8, 10, and 15 MV; 6‐ and 10‐MV flattening filter‐free), multileaf collimator (MLC) types (Millennium 120, High Definition 120, dual‐layer MLC), and disease sites (head and neck, pancreatic, and rectal cancer) and compare these parameters with those of clinical plans. Methods RapidPlan models in the Eclipse version 15.6 were used with the data of 28, 42, and 20 patients with head and neck, pancreatic, and rectal cancer, respectively. RapidPlan models of head and neck, pancreatic, and rectal cancer were created for TrueBeam STx (High Definition 120) with 6 MV, TrueBeam STx with 10‐MV flattening filter‐free, and Clinac iX (Millennium 120) with 15 MV, respectively. The models were used to create volumetric‐modulated arc therapy plans for a 10‐patient test dataset using all energy and MLC types at all disease sites. The Holm test was used to compare multiple dosimetric indices in different treatment machines and energy types. Results The dosimetric indices for planning target volume and organs at risk in RapidPlan model‐based plans were comparable to those in the clinical plan. Furthermore, no dose difference was observed among the RapidPlan models. The variability among RapidPlan models was consistent regardless of the treatment machines, MLC types, and energy. Conclusions Dosimetric indices of RapidPlan model‐based plans appear to be comparable to the ones based on clinical plans regardless of energies, MLC types, and disease sites. The results suggest that the RapidPlan model can generate treatment plans independent of the type of treatment machine.


| INTRODUCTION
To achieve clinical goals using volumetric-modulated arc therapy (VMAT), it is important to understand optimization methods and patient conditions. Thus, plan quality and optimization time of VMAT is dependent on the planners' knowledge and experience.
Knowledge-based VMAT planning was developed to minimize interplanner variability and improve plan quality. 1 Methods for knowledge-based planning can be further divided into two major categories: atlas-based methods and statistical modeling (including machine learning methods). 1 One of the statistical modeling methods is RapidPlan (Varian Medical Systems, Palo Alto, CA), a commercial knowledge-based planning solution derived from earlier work, which uses a model based on a library of previous plans. [2][3][4][5][6][7][8][9][10][11] The model can be used to predict a range of achievable organ at risk (OAR) dose-volume histograms (DVHs) for new patients. It is possible to share models among different clinical institutes in a cooperative framework. [12][13][14][15][16] Although knowledge-based planning using RapidPlan ensures efficiency in clinical practice, the model parameters in RapidPlan affect the quality of the predicted DVHs. 17 The statistical information in RapidPlan models varies according to the plan and treatment machine parameters due to the geometry-based expected dose (GED) calculation step. 18

2.B | Contouring and treatment plan
All critical structures, such as OARs, were contoured by radiation oncologists and medical physicists. Target volumes were contoured by radiation oncologists. In addition, all plans were optimized by several expert radiation oncologists and medical physicists who were responsible for the protocol in clinical practice at the time of model generation. [22][23][24] The radiation dose calculation algorithm used for Eclipse was Acuros XB (dose-to-medium) with heterogeneity correction. The calculation grid size was 2.5 mm.
Radiotherapy treatment in simultaneous integrated boost VMAT of HNC patients was set to 70 Gy in 35 fractions using TrueBeam STx (Varian Medical Systems) with 6 MV. The gross tumor volume (GTV) was defined as the gross extent of tumor evident in computed tomography (CT) images, including both the primary tumor and gross regional lymph nodes. The clinical target volume (CTV) was defined as the GTV plus a margin allowing for potential microscopic tumor extension and encompassing the adjacent regional lymph nodes. The PTV was the CTV plus a 5-mm-wide margin to allow for uncertainties in radiation delivery and the internal and set-up margins. The GTV, CTV, and PTV were defined according to the contouring policy described in a previous report. 22 The PTV70 volume included the primary tumor and lymph node metastases, whereas PTV63 and PTV56 volumes included high-risk and low-risk lymph nodes, respectively. The spinal cord and the left and right parotid glands were evaluated as OARs. The prescription dose was specified as D 50% (the dose that covers 50% of the structure) to PTV. The dose constraints are shown in the supporting information Table S1.
In PK patients, treatment prescription was set to 45 Gy in 15 fractions using TrueBeam STx with 10-MV FFF. The target delineation, including GTV, CTV, and PTV, is described in the study by Goto et al. 23 The prescription dose was specified as D 95% to PTV.
The spinal cord, stomach, and duodenum were defined as OARs. The dose constraints for OARs based on a previous institutional trial are shown in supporting information Table S2.

2.C | RapidPlan model creation
The overall study scheme is shown in Fig. 1 which were tuned to achieve our institution's acceptance criteria. The prescription setting was the same as for the clinical treatment plan.
The RapidPlan templates for planning optimization of all disease sites are shown in supporting information Tables S4-S6.

2.D | Dosimetric assessment
The models were used to create VMAT plans for a 10-patient test dataset using an open-loop process. 6 The treatment machine information is shown in supporting information Table S7. The RapidPlan model was calculated in this test dataset for all energy and MLC types at all disease sites. In the HNC group, the 6-MV and 6-MV FFF energy types were only used because other high-energy parameters are inadequate in clinical practice due to the lower skin dose.
All RapidPlan models were compared in terms of PTV and OAR parameters against original clinical plans. Selected relevant dose statistics and dose-volume parameters were considered. Concerning PTVs, D 98% and D 2% were considered. Regarding OARs, the following parameters were assessed: D max in the spinal cord and D mean in the parotids for HNC, D max in the spinal cord, as well as V 36Gy and V 39Gy in the stomach and duodenum for PK, and V 15Gy in the large and small bowels for RC. | 247 types to assess the statistically significant differences. A statistically significant difference was defined as P < 0.05.

| RESULTS
The heterogeneity of the test dataset was appraised in terms of the variability of PTV and main OAR volumes for each disease site. Supporting information Table S8 shows the comparison between the training and test datasets for PTV and OAR volumes. Only the difference in duodenum volumes between the training and test datasets was statistically significant (P = 0.02).
Example dose distribution of both clinical plan and RapidPlan model-based plan in HNC, PC, and RC is shown in Fig. 2. The mean ± standard deviation (SD) of the dosimetric indices in the HNC group is shown in Table 1 For PK, the mean ± SD of the dosimetric indices is summarized in Table 2  TrueBeam STx with HD, and for Halcyon in this study. To the best of our knowledge, the differences between the clinical plan and the RapidPlan model-based plan have not yet been compared for these treatment machines. We hope that RapidPlan has the potential to eliminate quality disparities not only regarding interplanner variability but also for generating consistent treatment plans among different treatment machines.
In HNC patients, 6 MV or 6-MV FFF was used with the test dataset to evaluate the MLC-type differences in these models. Hong et al. reported that MLC with a finer leaf width (2.5 mm) showed better dosimetric characteristics, providing better dose conformity to the target and reducing spinal cord and peripheral doses in HNC patients; however, no significant difference in dosimetric error was observed according to the MLC leaf width. 25 Li et al. compared TrueBeam of the Millennium 120 MLC with the Halcyon dual-layer MLC. 26 They described that the MLC width may still have an effect on normal tissue doses, although statistical differences were not observed. 26 Consequently, the effect of MLC width on dosimetric deviations in HNC VMAT plans was small. The dose distributions at PTVs and OARs created by the RapidPlan model in our test dataset were similar to those of the prior study. 26 In RapidPlan, the dose distribution in the OAR is partitioned into four regions: overlap region between PTV and OAR, in-field region, MLC transmission region, and out-of-field region. These four regions were used to calculate the GED and predict the new DVH. MLC-type differences affected the dose distribution in the MLC transmission region, although this region was smaller than other regions. 17 Thus, the effect of the MLC type on dose distribution in the RapidPlan model-based plan was associated with low-to-middle dose distributions, in particular to OARs.
In PK cases, the effect of energy types on the differences between clinical plans and RapidPlan model-based plans was small.   Abbreviations: PTV, planning target volume; FFF, flattening filter free; D xx% , dose covering xx% volume of region of structure; V yyGy , volume receiving yy Gy.
The mean ± standard deviation of dosimetric indices for rectal cancer. participated in its design and coordination, and helped to draft the manuscript. All authors read and approved the final manuscript.

SUPPORTING INFORMATION
Additional supporting information may be found online in the Supporting Information section at the end of the article.