Abstract
Background
The aim of this study was to clarify factors predicting the performance of knowledge-based planning (KBP) models in volume modulated arc therapy for prostate cancer in terms of sparing the organ at risk (OAR).
Materials and methods
In three institutions, each KBP model was trained by more than 20 library plans (LP) per model. To validate the characterization of each KBP model, 45 validation plans (VP) were calculated by the KBP system. The ratios of overlap between the OAR volume and the planning target volume (PTV) to the whole organ volume (Voverlap/Vwhole) were analyzed for each LP and VP. Regression lines between dose–volume parameters (V90, V75, and V50) and Voverlap/Vwhole were evaluated. The mean OAR dose, V90, V75, and V50 of LP did not necessarily match those of VP.
Results
In both the rectum and bladder, the dose–volume parameters for VP were strongly correlated with Voverlap/Vwhole at institutes A, B, and C (R > 0.74, 0.85, and 0.56, respectively). Except in the rectum at institute B, the slopes of the regression lines for LP corresponded to those for VP. For dose–volume parameters for the rectum, the ratios of slopes of the regression lines in VP to those in LP ranged 0.51–1.26. In the bladder, most ratios were less than 1.0 (mean: 0.77).
Conclusion
For each OAR, each model made distinct dosimetric characterizations in terms of Voverlap/Vwhole. The relationship between dose–volume parameters and Voverlap/Vwhole of OARs in LP predicts the KBP models’ performance sparing OARs.
Keywords: Knowledge-based planning, RapidPlan, Volumetric modulated arc therapy, Inverse planning, Quality assurance for planning
1. Introduction
Knowledge-based planning (KBP) was designed to compensate for the weaknesses of inverse planning, in which plan quality depends on planners or institutional skills.1, 2, 3, 4, 5, 6, 7 KBP can provide dose distributions for new patients according to the relationships, dose distributions, and displacements of the target and organs at risk (OARs) at an institution. Formulas to estimate the dose to OARs were derived from the relationship between dose and displacement.1, 2, 3 Moore et al.3 described KBP as the control of dosimetric variations among calculated plans.
The KBP system named RapidPlan (Varian Medical Systems, Palo Alto, CA, USA) was developed and is utilized in a commercial treatment planning system, Eclipse (Varian). Various deep learning and machine learning (ML) techniques have spread in radiology.8, 9, 10, 11, 12 RapidPlan’s ML technique is a supervised model for calculation that is trained with more than 20 cases. In RapidPlan, the objectives are set automatically according to learning plans in the models.12 Using RapidPlan has provided more stable OAR doses with shorter planning time than manual planning.7,13, 14, 15, 16, 17
RapidPlan has the advantage of sharing its models among multiple institutions. It has been shown that sharing models between institutions with similar planning designs effectively facilitated standardization of inverse planning.18 Because performance of KBP models depends on the model’s plan design,19, 20 when planners share KBP models from other institutions with a superior plan design, planners may improve their institutional planning with the KBP system. Thus, planners should test whether the models are suitable for their institutional plan design at least once before sharing them. In addition, verifying the performance of KBP models takes much time because calculating plans for various cases with KBP is required. Thus, predictions of KBP model performance without calculation are necessary. The aim of this study was to clarify factors predicting the performance of KBP models in terms of sparing the OAR. In KBP using RapidPlan, the method of predicting the dose to the OAR with plans being used to train the models was suggested.
2. Methods
2.1. Definitions of structures and prescriptions at each institute
Three institutions (Institutes A, B, and C) were enrolled in this study. These institutes treated patients with T1–T2c prostate cancer using volumetric modulated arc therapy (VMAT). Each institute has different contouring definitions for the clinical target volume (CTV), CTV to the planning target volume (PTV) margin, and rectum. Contouring definitions for the CTV were prostate and proximal 10 mm of seminal vesicle (SV), prostate and proximal half of SV, and prostate and proximal 15 mm of SV in institute A, B, and C. CTV to PTV margin in mm for all around, except posteriorly/posteriorly, are 6/4, 8/6, and 8/5 in institute A, B, and C. Contouring definitions for the rectum are from the rectosigmoid junction to the anus, 1.5 cm above the SV to 1.5 cm below the prostate, and up to 1.0 cm above and below the PTV in institute A, B, and C. The prescription dose was a mean PTV dose of 78 Gy at Institutes A and B and the dose to a volume including 95% (D95) of the PTV minus the rectum at Institute C.
2.2. Each RapidPlan model for calculation of KBP
At each institute, KBP models were created using more than 20 library plans (LP) for each model in the RapidPlan system (Varian Medical Systems, Palo Alto, CA, USA). In each model, LP were clinically used for treatment before April 2017 at each institute. To attain the ideal dose distribution in RapidPlan, the upper objectives were set to decrease the high dose region in the rectum for each institute. For the line objectives, the mean ± SD of priority was 64 ± 0.7, 60 ± 1.6, and 49 ± 2.3 at Institutes A, B, and C, respectively. Each clinically accepted model from all institutes was sent to institute B and imported into the Eclipse there. These models were read from. xml files uploaded to the Varian Model Analytics (https://ModelAnalytics.varian.com) website. The files also contained basic information about the model’s LP, such as dose volume histogram (DVH) and the ratio of overlap between OAR volume and PTV to the whole organ volume (Voverlap/Vwhole).
2.3. Validation plans for KBP
To validate the performance of 3 kinds of KBP models, 45 cases clinically treated from May 2017 to April 2018 were used with single optimization in Institute B’s RapidPlan system. The validation plans (VP) were different from the LP in each model. The CT scanner was Revolution HD (GE Medical Systems, Milwaukee, WI, USA). The slice thickness of the CT images was 2 mm, matrix size 512 × 512, and field of view 50 cm.
For these VPs, the following beam conditions were utilized: beam energy 6 MV photons from a TrueBeam linear accelerator equipped with a Millennium 120-leaf multileaf collimator (Varian Medical Systems, Palo Alto CA, USA). The treatment field was 2 full arcs of clockwise rotation from 181˚ to 179˚ and counterclockwise from 179˚ to 181˚. The arcs had collimator rotation of ±30˚, and the optimization and calculation algorithms were Anisotropic Analytical Algorithm and Photon Optimizer 13.0 (Varian Medical Systems, Palo Alto CA, USA). The grid size was 2.5 mm. The prescription setting for VP was the same as that for LP in each institution. Written informed consent was obtained from all patients, and the Institutional Ethics Committee approved this study (Osaka International Cancer Institute Review Board number: 1611119172).
2.4. Data analysis
The dose–volume parameters, represented as a percentage of the prescribed dose to 2%, 95%, and 98% of the prescribed volume (D95, D95, and D98) and the volume ratio receiving 50%, 75%, and 90% of the prescribed dose (V50, V75, and V90, respectively) for the rectum and bladder, were extracted from the DVH data for LP and VP. To evaluate the OAR dose of each model, each dose–volume parameter was sorted by Voverlap/Vwhole in the rectum and bladder. The correlations between dose–volume parameters (V50, V75, and V90) and Voverlap/Vwhole were evaluated using linear regression analysis. The Mann-Whitney U test (SPSS 8.0; SPSS, Inc., Chicago, IL) was used to calculate and evaluate the differences between the dose–volume parameters in LP and VP. Values of p < 0.05 were defined as significant.
3. Results
3.1. Basic information about cases in the model and validation sets
Table 1 shows basic information about the LP and VP. The model at institute B was trained by the maximum number of plans (i.e., 100 plans). Institute A had the smallest average Voverlap/Vwhole for the rectum and bladder among the three institutions, although the average volume of the rectum and bladder were also the smallest at Institute A. The use of the smallest margin setting at Institute A might have influenced this result. Institute B had the largest Voverlap/Vwhole ratios for the rectum and bladder among the three institutions because institute B had the largest PTV margin.
Table 1.
Information about each model. Numbers of cases, whole organ volume [cm3], and ratio of overlap between organ at risk (OAR) and planning target volume to the whole organ volume (Voverlap/Vwhole) for library plans (LP) and validation plans (VP) using each model.
| OAR | LP |
VP | |||
|---|---|---|---|---|---|
| Institute | A | B | C | ||
| Case number | 20 | 100 | 50 | 45 | |
| Rectum | Total volume [cm3] | 42.4 ± 13.0 | 45.2 ± 14.1 | 59.7 ± 24.9 | 44.2 ± 15.4 |
| Voverlap/Vwhole | 5.8% ± 2.2% | 8.8% ± 3.6% | 6.0% ± 2.7% | 11.3% ± 6.3% | |
| Bladder | Total volume [cm3] | 98.9 ± 26.8 | 173.8 ± 101.7 | 165.1 ± 98.4 | 160.2 ± 79.4 |
| Voverlap/Vwhole | 5.1% ± 2.3% | 8.3% ± 4.5% | 8.7% ± 3.9% | 8.8% ± 6.0% | |
Table 2 shows dosimetric comparison of LP and VP in each institution’s model. In the PTV, there were significant differences between VP and LP, except D95 for institute C. In the rectum, at institute A, each dose–volume parameter for VP was significantly larger than that for LP (p < 0.01). Institute B and C have significant differences in V50 between VP and LP (p < 0.01). In the bladder, at institute A, each dose–volume parameter for VP was significantly larger than that for LP (p < 0.05). Institute B had no significant differences in any dose–volume parameters between VP and LP. Institute C had significant differences in V75 and V90 between VP and LP (p < 0.05).
Table 2.
Dosimetric comparison of library plans (LP) and validation plans (VP) in each institution’s model.
| A | B | C | |||||
|---|---|---|---|---|---|---|---|
| LP | VP | LP | VP | LP | VP | ||
| PTV | D2 | 103.0 ± 0.3 | 102.1 ± 0.5 | 103.4 ± 0.4 | 102.4 ± 0.6 | 105.6 ± 0.8 | 103.7 ± 0.9 |
| p val. | <0.01 | <0.01 | <0.01 | ||||
| D95 | 96.4 ± 0.6 | 96.9 ± 0.5 | 96.2 ± 0.3 | 96.8 ± 0.7 | 100 ± 0.4 | 100 ± 0.0 | |
| p val. | <0.01 | <0.01 | 0.195 | ||||
| D98 | 93.7 ± 1.3 | 94.7 ± 1.0 | 92.9 ± 1.5 | 94.8 ± 1.0 | 98.8 ± 0.7 | 99.3 ± 0.3 | |
| p val. | 0.03 | <0.01 | <0.01 | ||||
| Rectum | V50 | 21.5 ± 5.5 | 28.6 ± 10.3 | 43.6 ± 6.7 | 40.6 ± 7.0 | 33.0 ± 7.6 | 36.7 ± 6.7 |
| p val. | <0.01 | <0.01 | <0.01 | ||||
| V75 | 10.9 ± 3.8 | 19.6 ± 8.9 | 23.9 ± 4.2 | 24.0 ± 7.6 | 15.9 ± 4.0 | 17.8 ± 4.4 | |
| p val. | <0.01 | 0.625 | 0.052 | ||||
| V90 | 4.4 ± 1.8 | 9.5 ± 6.2 | 14.1 ± 3.0 | 12.9 ± 5.3 | 7.9 ± 2.6 | 8.7 ± 2.6 | |
| p val. | <0.01 | 0.372 | 0.050 | ||||
| Bladder | V50 | 24.2 ± 8.6 | 34.1 ± 15.0 | 33.7 ± 14.2 | 29.2 ± 14.3 | 41.7 ± 14.8 | 37.6 ± 16.5 |
| p val. | 0.03 | 0.07 | 0.188 | ||||
| V75 | 13.6 ± 5.1 | 18.2 ± 8.5 | 16.9 ± 7.6 | 16.4 ± 9.5 | 23.5 ± 8.5 | 19.9 ± 9.7 | |
| p val. | 0.017 | 0.395 | 0.040 | ||||
| V90 | 8.9 ± 3.4 | 12.2 ± 6.2 | 10.9 ± 5.0 | 11.3 ± 7.1 | 16.2 ± 5.9 | 13.8 ± 7.3 | |
| p val. | 0.013 | 0.778 | 0.043 | ||||
Abbreviations:PTV planning target volume, V50 volume ratio receiving 50% of the prescribed dose, V75 volume ratio receiving 75% of the prescribed dose, V90 volume ratio receiving 90% of the prescribed dose, D2 percentage of the prescribed dose to 2% of the prescribed volume, D95 percentage of the prescribed dose to 95% of the prescribed volume, D98 percentage of the prescribed dose to 98% of the prescribed volume.
3.2. Relationships between dose–volume parameters and Voverlap/Vwhole
Fig. 1 shows the relationships between V75 and Voverlap/Vwhole for the rectum at institutes A, B, and C. Linear regression lines were drawn between dose–volume parameters and Voverlap/Vwhole. For V75 at institutes A and C, the slope and height of VP regression lines coincided with those for LP. For V75 at institute B, the height of the VP regression lines was lower than those for LP, although the slope of the regression lines for VP coincided with that for LP. The regression lines for VP ran along the lower boundary of the LP distributions. The manual plans used clinically for VP are shown in Fig. 1 (b). For manual plans in which the rectal Voverlap/Vwhole ratios were large, the rectal dose was reduced so as not to exceed the dose constraints. Therefore, the rectal doses in the manual plans were lower than those of the VP optimized by RapidPlan.
Fig. 1.
Relationship between V75 and Voverlap/Vwhole of the rectum at institutes A (a), B (b), and C (c). The horizontal axis is Voverlap/Vwhole for the rectum. The vertical axis is V75 for the rectum. Blue and orange dots represent the doses in library plans (LP) and validated plans (VP), respectively. Each solid line is a linear regression line between each rectal dose and Voverlap/Vwhole for the rectum. Gray dots represent the doses in manual plans for validated cases.
Fig. 2 shows the relationships between V75 and Voverlap/Vwhole for the bladder at institutes A, B, and C. At all institutes, the height of the VP regression lines coincided with those for LP. Most of the slopes of the regression lines for VP were smaller than those for LP. The intervals between the regression lines of VP and LP were spread so that the range of Voverlap/Vwhole was large. Manual plans for VP are shown in Fig. 2 (b). The bladder dose in the manual plans was consistent with that for VP in all cases.
Fig. 2.
Relationship between V75 and Voverlap/Vwhole of the bladder at institutes A (a), B (b), and C (c). The horizontal axis is Voverlap/Vwhole for the bladder. The vertical axis is V75 for the bladder. Blue and orange dots represent the doses in library plans (LP) and validated plans (VP), respectively. Each solid line is a linear regression line between each bladder dose and Voverlap/Vwhole for the bladder. Gray dots represent the doses in manual plans for validated cases.
Table 3 shows the correlation values between all dosimetric parameters and Voverlap/Vwhole for each OAR in LP and VP for each KBP model. In the rectum, VP for all dosimetric values, except V90 at institute A, had higher correlation values than those for LP. At Institutes A and B, all dosimetric values for VP had strong correlation values (R ≥ 0.9). In the bladder, each dose–volume parameter for VP and LP had strong correlations with Voverlap/Vwhole for each OAR (R > 0.7).
Table 3.
Correlation values between Voverlap/Vwhole and each dosimetric value for library plans (LP) and validation plans (VP) using each model.
| V50 | V75 | V90 | |||||
|---|---|---|---|---|---|---|---|
| OAR | Institute | LP | VP | LP | VP | LP | VP |
| Rectum | A | 0.76 | 0.93 | 0.89 | 0.97 | 0.99 | 0.98 |
| B | 0.41 | 0.96 | 0.78 | 0.98 | 0.78 | 0.99 | |
| C | 0.48 | 0.57 | 0.50 | 0.73 | 0.34 | 0.63 | |
| Bladder | A | 0.78 | 0.74 | 0.91 | 0.90 | 0.91 | 0.94 |
| B | 0.87 | 0.85 | 0.97 | 0.96 | 0.99 | 0.99 | |
| C | 0.83 | 0.76 | 0.88 | 0.93 | 0.90 | 0.97 |
Abbreviations:Voverlap/Vwhole ratio of overlap between OAR volume and planning target volume to the whole organ volume, V50 volume ratio receiving 50% of the prescribed dose, V75 volume ratio receiving 75% of the prescribed dose, V90 volume ratio receiving 90% of the prescribed dose.
Table 4 shows the ratios of slopes of regression lines between all dose–volume parameters and Voverlap/Vwhole for each OAR in LP and VP for each KBP model. The slopes of the regression lines represent the performance of increasing the OAR dose per unit Voverlap/Vwhole in each KBP model. Thus, smaller values were superior at sparing the OAR dose per unit Voverlap/Vwhole. The ratio lower than 1.0 means that sparing the OAR dose per unit Voverlap/Vwhole for VP is superior to that for LP. In the rectum, the ratio of the slope in Table 4 ranged 0.51–1.26 (mean: 0.95; ratio >1.0 means higher slopes for VP than LP). In V75 for the rectum, sparing the OAR dose per unit Voverlap/Vwhole for VP is improved compared with that for LP in each institute’s model.
Table 4.
The rata of slopes of linear regression lines between Voverlap/Vwhole and each dosimetric value for library plans (LP) and validation plans (VP) using each model.
| Institute | V50 | V75 | V90 | |
|---|---|---|---|---|
| Rectum | A | 0.83 | 0.92 | 1.26 |
| B | 1.16 | 0.94 | 1.01 | |
| C | 0.51 | 0.79 | 0.91 | |
| Bladder | A | 0.63 | 0.64 | 0.71 |
| B | 0.70 | 0.88 | 1.04 | |
| C | 0.67 | 0.79 | 0.87 |
Abbreviations:Voverlap/Vwhole ratio of overlap between OAR volume and planning target volume to the whole organ volume, OAR organ at risk, V50 volume ratio receiving 50% of the prescribed dose, V75 volume ratio receiving 75% of the prescribed dose, V90 volume ratio receiving 90% of the prescribed dose.
4. Discussion
In this study, characterizations of LP and VP at three institutions with different plan designs and contouring definitions were analyzed in terms of the Voverlap/Vwhole of two OARs: the rectum and bladder. Strong correlations between Voverlap/Vwhole and OAR dose were observed in three KBP models. The strong correlation between Voverlap/Vwhole and OAR dose in KBP has already been described3,19; however, few reports have mentioned the relationship between LP and VP in KBP. The regression lines between dosimetric parameters and Voverlap/Vwhole of the rectum in LP coincided with the regression lines in VP, except those for the rectum at institute B.
The previous report described that estimated OAR dose in LP had strong correlation with Voverlap/Vwhole in multiple centers.19 They could not create the relationships between OAR dose calculated with RapidPlan and Voverlap/Vwhole in VP because there were two VPs. In this study, 45 plans were calculated with RapidPlan for validations in each model. Therefore, we can compare regression curves between OAR dose and Voverlap/Vwhole in LP and VP. We found that the relationship between dose–volume parameters and Voverlap/Vwhole of OARs in LP predicted the KBP models’ performance without dependencies on the number of LP and range of Voverlap/Vwhole. In contrast, the manual plans did not show correlations between dose–volume parameters and Voverlap/Vwhole, as shown in Fig. 1, Fig. 2. Therefore, this methodology can be applied for all users. The results of this study indicated that users should select plans to train models with dosimetric analysis when they create KBP models.
The averages of each dosimetric parameter in VP were inconsistent with those in LP in Table 2. Generally, in the cases which Voverlap/Vwhole is small, it is easy to reduce organs’ dose. As shown in Table 1, the organs’ volume and Voverlap/Vwhole in LP is different from those for VP at each institute. Therefore, the averages of each dosimetric parameter in VP were inconsistent with those in LP. The OAR dose in KBP depended on the Voverlap/Vwhole for the OAR in each model, we suggest selecting optimal KBP models for each case. According to the formulas of LP and VP in Fig. 1, Voverlap/Vwhole for the rectum was over 20%, and V75 of the rectum was over 30% in the models at Institutes A and B. Thus, when a V75 value of less than 20% is required for cases in which Voverlap/Vwhole for the rectum is 20%, the model created at Institute C is effective. According to the regression line formulas for VP in Fig. 1, when Voverlap/Vwhole for the rectum was 5%, the V75 values for LP and VP at Institute A were the smallest among the three institutes. Therefore, in cases with a Voverlap/Vwhole value of 5% for the rectum, the model created at Institute A is effective. To select the optimal model from multiple models, characterization of OAR dose for LP in terms of Voverlap/Vwhole for the OAR should be evaluated.
Wang et al.21 suggested a method to evolve KBP models in which plans with lower doses to OARs were re-registered to improve the models and reduce the doses to the OARs by several percent. The results of this study indicate that differences in LP greatly impact VP. The analysis of V75 for the rectum in VP at Institutes A and B (Fig. 1(b)) showed that when Voverlap/Vwhole was 10%, the rectal doses calculated by each regression formula were 16.9% and 22.3% at institutes A and B, respectively, although there were small differences between the PTV values of the two institutions. To improve the KBP model, plans with lower doses to OARs (instead of the original plans) should be re-registered to the model.
Moreover, we investigated whether LP could predict the results for KBP even if the model was trained by the minimum number of cases in LP.19 Fig. 1, Fig. 2 show that at institution A, which had the minimum number of 20 cases, the regression lines for LP coincide with those for VP in comparison with the results at other institutions. This indicates that even if the number of LP is 20, the calculation results of KBP can be predicted by LP, i.e., the number of the cases used to train the model was independent of the KBP prediction.
There are two main OARs for the prostate: the rectum and bladder. These two OARs have different limitations to calculate the dose distribution in inverse planning. Planners should focus on the high and medium dose regions of the rectum.22 This study employed three institutions’ original plan designs for the rectum. The plan design of Institute C reduced the dependence of rectal dose on Voverlap/Vwhole; therefore, the slopes of its regression lines for LP (Fig. 1) were the smallest among the three institutions. The plan design affected characterizations for VP, so institute C’s regression slopes of VP were also the smallest among the three institutions. In contrast, the relationships of the regression lines for all dose–volume parameters between LP and VP were similar among the three institutions. In VP, the variation of the slopes for the bladder (Fig. 2 and Table 3) was smaller than that for the rectum across all three institutions.
Chang et al. set the upper objectives manually to improve KBP with RapidPlan.23,24 In this study, the upper objectives (except line objectives) were set automatically at each institution to decrease high-dose regions in the rectum because the maximum dose is more important in serial organs such as the rectum. The upper objectives were decided at each institute to suit each institute’s dose constraints. When the models created by different institutes are shared, it is of particular concern that upper objectives be set according to various institutional constraints. In RapidPlan, the line objectives are placed horizontally in the overlap region between the PTV and OARs to prevent underdosing the PTV,25 which increases the high-dose volume.7 These results indicate that some objectives (but not line objectives) are required to attain ideal dose distributions with RapidPlan. To reduce dose variation between planners, the RapidPlan system should be improved to attain ideal dose distributions with only line objectives.
5. Conclusion
Dosimetric characterizations of the rectum and bladder using LP and VP in KBP were analyzed. Performance in terms of reducing dosimetric variations to OARs depends on each model. Sorting doses to OARs in LP in terms of Voverlap/Vwhole was an efficient method to predict the performance of VP.
Conflict of interest
None declared.
Financial disclosure
This study was supported by JSPS KAKENHI Grant (17K15817) to Yoshihiro Ueda. The authors gratefully acknowledge the support provided by the Japanese Society of Radiological Technology (JSRT) Research Grant (2019, 2020).
Acknowledgements
We thank Richard Lipkin, PhD, from Edanz Group (www.edanzediting.com/ac) for editing a draft of this manuscript.
References
- 1.Hunt M.A., Jackson A., Narayana A., Lee N. Geometric factors influencing dosimetric sparing of the parotid glands using IMRT. Int J Radiat Oncol Biol Phys. 2006;66:296–304. doi: 10.1016/j.ijrobp.2006.05.028. [DOI] [PubMed] [Google Scholar]
- 2.Wu B., Ricchetti F., Sanguineti G. Patient geometry-driven information retrieval for IMRT treatment plan quality control. Med Phys. 2009;36(12):5497–5505. doi: 10.1118/1.3253464. [DOI] [PubMed] [Google Scholar]
- 3.Moore K.L., Brame R.S., Low D.A., Mutic S. Experience-based quality control of clinical intensity-modulated radiotherapy planning. Int J Radiat Oncol Biol Phys. 2011;81(2):545–551. doi: 10.1016/j.ijrobp.2010.11.030. [DOI] [PubMed] [Google Scholar]
- 4.Williams M.J., Bailey M.J., Forstner D., Metcalfe P.E. Multicentre quality assurance of intensity-modulated radiation therapy plans: A precursor to clinical trials. Australas Radiol. 2007;51(5):472–479. doi: 10.1111/j.1440-1673.2007.01873.x. [DOI] [PubMed] [Google Scholar]
- 5.Chung H.T., Lee B., Park E., Lu J.J., Xia P. Can all centers plan intensity-modulated radiotherapy (IMRT) effectively? An external audit of dosimetric comparisons between three-dimensional conformal radiotherapy and IMRT for adjuvant chemoradiation for gastric cancer. Int J Radiat Oncol Biol Phys. 2008;71(4):1167–1174. doi: 10.1016/j.ijrobp.2007.11.040. [DOI] [PubMed] [Google Scholar]
- 6.Bohsung J., Gillis S., Arrans R. IMRT treatment planning—A comparative inter-system and inter-Centre planning exercise of the QUASIMODO group. Radiother Oncol. 2005;76:354–361. doi: 10.1016/j.radonc.2005.08.003. [DOI] [PubMed] [Google Scholar]
- 7.Kubo K., Monzen H., Ishii K. Dosimetric comparison of RapidPlan and manually optimized plans in volumetric modulated arc therapy for prostate cancer. Phys Med. 2017;S1120-1790(17):30216–30218. doi: 10.1016/j.ejmp.2017.06.026. [DOI] [PubMed] [Google Scholar]
- 8.Lustberg T., van Soest J. Clinical evaluation of atlas and deep learning based automatic contouring for lung cancer. Radiother Oncol. 2018;126(Feb (2)):312–317. doi: 10.1016/j.radonc.2017.11.012. [DOI] [PubMed] [Google Scholar]
- 9.Cardenas C.E., McCarroll R.E., Court L.E. Deep learning algorithm for auto-delineation of high-risk oropharyngeal clinical target volumes with built-in dice similarity coefficient parameter optimization function. Int J Radiat Oncol Biol Phys. 2018;101(June (2)):468–478. doi: 10.1016/j.ijrobp.2018.01.114. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Tong N., Gou S., Yang S., Ruan D., Sheng K. Fully automatic multi-organ segmentation for head and neck cancer radiotherapy using shape representation model constrained fully convolutional neural networks. Med Phys. 2018;(August) doi: 10.1002/mp.13147. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Yoon D.K., Jung J.Y., Jo Hong K. GPU-based prompt gamma ray imaging from boron neutron capture therapy. Med Phys. 2015;42(Jan (1)):165–169. doi: 10.1118/1.4903265. [DOI] [PubMed] [Google Scholar]
- 12.Yuan L., Ge Y., Lee W.R., Yin F.F., Kirkpatrick J.P., Wu Q.J. Quantitative analysis of the factors which affect the interpatient organ-at-risk dose sparing variation in IMRT plans. Med Phys. 2012;39(11):6868–6878. doi: 10.1118/1.4757927. [DOI] [PubMed] [Google Scholar]
- 13.Tol J.P., Delaney A.R., Dahele M., Slotman B.J., Verbakel W.F. Evaluation of a knowledge-based planning solution for head and neck cancer. Int J Radiat Oncol Biol Phys. 2015;91(3):612–620. doi: 10.1016/j.ijrobp.2014.11.014. [DOI] [PubMed] [Google Scholar]
- 14.Fogliata A., Nicolini G., Clivio A. A broad scope knowledge-based model for optimization of VMAT in esophageal cancer: Validation and assessment of plan quality among different treatment centers. Radiat Oncol. 2015;10(1):220. doi: 10.1186/s13014-015-0530-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Wu H., Jiang F., Yue H., Li S., Zhang Y. A dosimetric evaluation of knowledge-based VMAT planning with simultaneous integrated boosting for rectal cancer patients. J Appl Clin Med Phys. 2016;17(6):78–85. doi: 10.1120/jacmp.v17i6.6410. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Scaggion A., Fusella M., Roggio A. Reducing inter- and intra-planner variability in radiotherapy plan output with a commercial knowledge-based planning solution. Phys Med. 2018;53:86–93. doi: 10.1016/j.ejmp.2018.08.016. [DOI] [PubMed] [Google Scholar]
- 17.Tamura M., Monzen H., Matsumoto K. Mechanical performance of a commercial knowledge-based VMAT planning for prostate cancer. Radiat Oncol. 2018;13(Aug (1)):163. doi: 10.1186/s13014-018-1114-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Schubert C., Waletzko O., Weiss C. Intercenter validation of a knowledge based model for automated planning of volumetric modulated arc therapy for prostate cancer. The experience of the German RapidPlan Consortium. PLoS One. 2017;12(May (5)):e0178034. doi: 10.1371/journal.pone.0178034. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Ueda Y., Fukunaga J., Kamima T., Adachi Y., Nakamatsu K., Monzen H. Evaluation of multiple institutions’ models for knowledge-based planning of volumetric modulated arc therapy (VMAT) for prostate cancer. Radiat Oncol. 2018;46:46. doi: 10.1186/s13014-018-0994-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Monzen H, Tamura M, Ueda Y, et al. Dosimetric evaluation with knowledge-based planning created at different periods in volumetric-modulated arc therapy for prostate cancer: a multi-institution study. Radiol Phys Technol. in press. [DOI] [PubMed]
- 21.Wang M., Li S., Huang Y. An interactive plan and model evolution method for knowledge-based pelvic VMAT planning. J Appl Clin Med Phys. 2018;19(5):491–498. doi: 10.1002/acm2.12403. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Peeters S.T., Lebesque J.V., Heemsbergen W.D. Localized volume effects for late rectal and anal toxicity after radiotherapy for prostate cancer. Int J Radiat Oncol Biol Phys. 2006;64:1151–1161. doi: 10.1016/j.ijrobp.2005.10.002. [DOI] [PubMed] [Google Scholar]
- 23.Chang A.T.Y., Hung A.W.M., Cheung F.W.K. Comparison of planning quality and efficiency between conventional and knowledge-based algorithms in nasopharyngeal Cancer patients using intensity modulated radiation therapy. Int J Radiat Oncol Biol Phys. 2016;95(3):981–990. doi: 10.1016/j.ijrobp.2016.02.017. [DOI] [PubMed] [Google Scholar]
- 24.Kamima T., Ueda Y., Fukunaga J.I. Multi-institutional evaluation of knowledge-based planning performance of volumetric modulated arc therapy (VMAT) for head and neck cancer. Phys Med. 2019;64:174–181. doi: 10.1016/j.ejmp.2019.07.004. [DOI] [PubMed] [Google Scholar]
- 25.Tol J.P., Dahele M., Delaney A.R., Slotman B.J., Verbakel W.F. Can knowledge-based DVH predictions be used for automated, individualized quality assurance of radiotherapy treatment plans? Radiat Oncol. 2015;10:234. doi: 10.1186/s13014-015-0542-1. [DOI] [PMC free article] [PubMed] [Google Scholar]