MSc Physics and Astronomy Physics of Life and Health Master Thesis Knowledge-based radiotherapy treatment planning for stage III lung cancer patients by Evgenia Tourou 11128879 June 2018 60EC Supervisor/Examiner: Examiner: Wilko F.A.R. Verbakel, PhD Geert J. Streekstra, PhD Daily Supervisor: Alexander R. Delaney, MSc Radiotherapy Department, VUmc medical center
68
Embed
Knowledge-based radiotherapy treatment planning for stage ... · multiple field directions are used. Additionally, a multileaf collimator (MLC) is utilized to shape the radiation
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
MSc Physics and Astronomy
Physics of Life and Health
Master Thesis
Knowledge-based radiotherapy treatment planning for
stage III lung cancer patients
by
Evgenia Tourou
11128879
June 2018
60EC
Supervisor/Examiner: Examiner:
Wilko F.A.R. Verbakel, PhD Geert J. Streekstra, PhD
Daily Supervisor:
Alexander R. Delaney, MSc
Radiotherapy Department, VUmc medical center
ABSTRACT
Treatment of large volume lung cancer is carried out mostly using two techniques, the full-RapidArc
(f-RA) or the hybrid-RapidArc (h-RA). The choice between the two methods depends on the
individual characteristics of the patient, while the treatment planners often have to make both plans in
order to choose for the optimal treatment technique for the patient. However, manual treatment
planning is a labor-intensive and time consuming process which, in many cases, does not
yield consistent or optimal plans. RapidPlan (Varian Medical Systems, Palo Alto, USA), a
knowledge-based-planning solution, uses the dosimetry and geometry of previous treatment
plans to model and predict organ-at-risk (OAR) dose-volume-histograms (DVHs) for future
patients based solely on their geometry. The present study investigates the possibility of utilizing
RapidPlan, as a tool for selecting f-RA or h-RA technique for individual lung cancer patients, without
the requirement of creating actual treatment plans. A f-RA and a f-RA model were created,
consisting of 50 clinical plans each, and were used to generate dose predictions and
subsequently to optimize model-based plans (MBPs) for a group of 10 patients. MBPs quality
was analyzed by benchmarking MBPs against the manual plans (MPs) made by experienced
radiotherapy treatment planners. DVH prediction accuracy was analyzed by comparing predicted
vs achieved OAR dose metrics. Finally, the number of patients that would have been selected
for f-RA or h-RA based solely on OAR predictions was compared to the corresponding
number of patients that would have been selected based on the achieved OAR doses in
MBPs. MBPs improved contralateral lung (CL) and total lung (TL-PTV) mean dose compared to
the manual plans in both techniques. However, CL V5 in the f-RA MBPs increased compared
to the MPs. The target coverage was inferior in the MBPs compared to the MPs. RapidPlan
was able to accurately predict the mean dose of CL, but it consistently underestimated the
amount of sparing that could be achieved for TL-PTV. Based only on comparing single OAR
dose volumes, RapidPlan can accurately predict which technique gives the lower dose in 7-9
/10 cases. The results showed that RapidPlan is able to generate MBPs of comparable quality
to the MPs for f-RA and h-RA techniques, nevertheless, it requires further validation with a
more wise selection of priorities and using generated point-objectives instead of line
When observing the h-RA predictions for CL, the DVH that deviates most from the
predicted DVH-range is that of patient 3 (Appendix B). That plan has an unusual field set up
of the conventional fields. The oblique field is set to 140 degrees, while the field direction in
the model library is between 150 and 160 degrees for left-sided tumors (Table 2.5).
Therefore, a larger part of the lung is located in the beam direction, resulting in an unusually
high exiting dose for the CL that the model is not able to predict. This part of the lung is
already in the in-field region of the RapidArc component, thus the in-field region is the same
whether the conventional fields are used or not. Furthermore, the model takes into account
the final field set-up including the RapidArc component and only takes into account the
direction of the fields, not how many beams are there. Consequently, the conventional fields
do not play any role in the estimation algorithm. After excluding this plan, the correlation
coefficient between predicted and achieved Dmean increases to 0.90 and σ drops to 0.9 and
are now comparable to the figures for the f-RA model. The same happens for the V5
predictions; R2increaces to 0.90 and σ is 5.0%, while also the slope decreases to 1.05.
The MBP CL mean dose of the h-RA plan of patient 8 deviated from the predicted by
1.9Gy. For the RapidArc component of this plan, a full-arc with avoidance sector was used
containing a 250° irradiation arc. That caused the GED PC1 to be higher as explained before.
Full-arc was used also for patient 6. However, an additional oblique field was used on 160°,
even though the tumor is located in the right lung, causing direct irradiation to the CL. The
extra dose delivered by this field, compensated for the higher GED PC1.
43
TL-PTV
The dose of TL-PTV in the f-RA MBPs was generally increased in comparison to the
MPs. When looking at the DVH predictions (Appendix B) the objective line is placed higher
than the MP DVH-line in 7/10 cases (patients 1,2,3,4,7,9,10) in the low dose region, showing
that the model underestimated the amount of sparing that could be achieved.
A possible cause for the high predictions, despite that model statistics indicate that the
model training has been successful, is the presence of an outlier in the model that it may
overfit the data. There is a plan with high CD value (CD=26.8, while the threshold is 10), that
is away from the regression line but close to identity line in the residual plot. To further
investigate that, the cleaned model, from which the suggested outliers were removed (Section
3.1.2), was used to generate predictions for the TL-PTV. The predictions were again
consistently higher than the MPs. More specifically, in 8/10 patients the objective line was
placed above the manual TL-PTV DVH-line.
The largest deviation between the predicted and achieved TL-PTV DVH was
observed for patients 1 and 2. The TL-PTV structure of patient 2 was indicated as an outlier
by the f-RA model because of large volume (Table 2.1 and Table 2.6). The TL-PTV structure
of patient 1, although not indicated as an outlier by the DVH estimation algorithm, has
volume higher than mean+1sd of the model library structures. Consequently, it is possible
that the model dose not perform well for large structure volumes.
It should be noted that RapidPlan was able to pull the DVH-line much lower than the
optimization objective line for patient 1, 2 and 10 as can be seen in Appendix B. This shows
that the TL-PTV sparing during the optimization process was mainly driven by the CL lung
objective, as the CL volume accounts for around 60% of the TL-PTV volume. To confirm
that, the plan of patient 1 was optimized again, without the use of dose-volume objectives for
the TL-PTV. The resultant plan had only minor differences in mean doses: CL Dmean was
0.09 Gy lower, TL-PTV Dmean was 0.15Gy higher and IL Dmean increased by 0.55Gy, while
TL-PTV V5 and V20 remained the same.
3.5 Using predictions to select treatment technique
Achieved MBP f-RA dose parameters plotted against the achieved h-RA dose
parameter for CL, TL-PTV and ESO are shown in Figure 3.12. The cases where the
44
predicted f-RA parameter was lower than the predicted h-RA parameter (for ex. f-RA
Predicted CL Dmean < h-RA Predicted CL Dmean ) are marked in red.
Data points on the left side of the identity line show that the f-RA MBP gave lower
dose than the h-RA MBP, while data points highlighted with red indicate that the f-RA
method would result in lower dose based on the predictions. There are 8/10 correct
predictions for CL Dmean, and 7/10 for TL-PTV Dmean, 9/10 for TL-PTV V20, and 8/10 for
ESO Dmean.
Based on the MBPs results, the h-RA technique is almost always better for CL Dmean
and TL-PTV V20, and f-RA method is better for TL-PTV Dmean and ESO Dmean. However,
considering that the f-RA TL-PTV V20 predictions are consistenlty high as described above,
we need to take a closer look at the TL-PTV V20 resaults. When examining the results of the
MPs in Figure 3.13, f-RA MPs resaulted in lower TL-PTV V20 in 5/10 patients. In that case,
RapidPlan gave correct prediction for 7/10 cases.
45
Figure 3.12: MBP f-RA dose parameters plotted against the achieved h-RA dose parameter
for CL, TL-PTV and ESO. Marked in red are the cases where the predicted f-RA parameter
was lower than the predicted h-RA parameter. Solid line represents the identity line.
46
Figure 3.13: Manual f-RA TL-PTV V20 plotted against the manual h-RA TL-PTV V20.
Marked in red are the cases where the predicted f-RA TL-PTV V20 was lower than the
predicted h-RA TL-PTV V20.
47
4. DISCUSSION AND CONCLUSION
The purpose of this thesis was to evaluate the performance of the RapidPlan
knowledge-based planning solution in providing f-RA and h-RA plans for large volume lung
cancer patients. Additionally, it was examined whether RapidPlan predictions can serve as a
tool for selection of the optimal treatment technique for a new patient.
The results showed that RapidPlan can generate MBPs of comparable quality to the
MPs made by experienced treatment planners. The use of line-objectives throughout the
whole dose range improved CL and TL-PTV Dmean, while in some cases MBPs also reduced
the CL V5 and TL-PTV V20 for specific patients. The most notable difference between MBPs
and MPs for both techniques was the decreased PTV V95%. The use of line-objective for CL
and TL-PTV negatively affected the PTV coverage. On top of that, the use of somewhat
higher priorities for ESO and SC also prevented the dose to be equally distributed in the PTV,
reducing its homogeneity. Another possible reason for the lower homogeneity it might be the
use of NTO in the MBPs, which on the other hand led to increased conformity.
Another remarkable result is the increased CL V5 in the f-RA MBPs, caused by the
fact that manual optimization objectives are placed at very low dose in order to pull down the
DVH as much as possible. However, with a quite accurate model, the predictions will never
be low enough to generate objectives in a similar position to the manual ones. Since
increasing the priority of the whole objective-line would affect the PTV coverage, the only
solution seems to be the use of generated point-objectives with increased priorities, at the low
dose region. This could also solve the issue with low PTV V95%. This approach was tested in
three patients and showed a reduction in CL V5. RapidPlan does have a feature to
automatically generate the priorities. Ideally, the suitable priorities that ensure sufficient
OAR sparing while maintaining good homogeneity of the PTV should be generated by
RapidPlan itself. However, generated priorities still do not seem to work in this way.
Fogliata et al.26 have shown that RapidPlan VMAT plans improved IL Dmean, CL
Dmean, and CL V20. The last is contrary to the findings of the present study. However,
Fogliata et al. have used line-objective on IL instead of TL-PTV, while no objectives were
used around the CL V5 in the manual plans. On the other hand, at the VUmc reducing the CL
V5 as low as possible is of high clinical importance, and manual plans have very low V5.
Therefore is difficult to further reduce V5 in the MBPs. Furthermore, in Fogliata et al. paper,
48
RapidPlan improved PTV homogeneity. However, their study involved smaller PTV
volumes.
A limitation of this study was the problem in the GED calculation algorithm which is
incorrectly modeling the avoidance sector, and this could be a contributing reason to
inaccuracies and poor results. Because of this, the optimization of f-RA plans had to be done
twice, consuming more time as explained in the Appendix A. However, the problem is not
completely solved as it also exists in the model training algorithm. The avoidance sector
plans are considered as having beamlets in the avoidance sector directions, and therefore
some radiation is expected from these directions, while the partial-arc plans are correctly
modeled. This leads to inconsistencies in a model containing both avoidance sector plans and
partial-arc plans. The problem was communicated to a Varian representative who stated that
the bug will be corrected in the next version of RapidPlan.
RapidPlan was able to accurately predict the mean dose of CL, but it consistently
underestimated the amount of sparing that could be achieved for TL-PTV. After excluding
the possibility that this is caused by any outliers in the model, the reasoning may be found in
the modeling of TL-PTV structure. TL-PTV structure has a large variation in dose
distribution because CL is spared as much as possible whereas IL receives much more dose.
RapidPlan solves the problem of varying dose within a structure by using volume partitioning
and constructing distinct models for the parts of the organ that is out-of-field or overlapping
the target. However, in case there is large dose variation within one model structure, it is
possible that this is more difficult to present with the PC’s. This hypothesis is further
supported by the observation that predictions for the IL, although not used for the
optimization, are of good quality (Figure B5 in Appendix B). If this is the cause of the
problem, a possible solution could be to utilize the predictions of the individual lungs to be
sync to the total lung predictions. However, there is no such mechanism in RapidPlan.
Furthermore, it is possible that the problem with not modeling correctly the avoidance
sector causes the inaccuracy in TL-PTV predictions. The GED-PCS1 seems to have a more
important role in the TL-PTV predictions since its partial coefficient of determination is
0.837 for the TL-PTV, while for the CL and IL is considerably lower: 0.737 and 0.663
respectively.
The h-RA MBPs, apart from the decreased PTV V95%, were of very good quality, and
resembled well the MPs since 90% of the dose is delivered by the conventional fields, which
were the same as in the MPs. RapidPlan is not designed for hybrid techniques and thus does
not take into account the conventional component of the field. However, because the model
49
consisted mostly of fields having a similar arrangement, the model incorporated correctly the
dosimetry of the training set, and thus generated quite accurate predictions for CL and TL-
PTV Dmean. On the contrary, the model was not able to predict correct dose in plans with
different field set-up than the standard. Nevertheless, the h-RA predictions for ESO were
lower than the achieved MBPs in 7/10 patients. That is because the esophagus is a small
structure, compared to the PTV, and the MLC movement and dose modulation in VMAT, can
sufficiently reduce its dose.
RapidPlan can better predict CL and TL-PTV mean doses, instead of dose-specific
volumes such as V5 and V20. When it comes to comparison between f-RA and h-RA, based
on the simple method used in this study, by only comparing single OAR dose volumes,
RapidPlan can accurately predict which technique gives the lower dose in 7-9 /10 cases.
To conclude, this study demonstrated that RapidPlan is capable of generating
clinically acceptable f-RA and h-RA plans for lung cancer patients. However, the plans could
be improved after a more wise selection of priorities and using generated point-objectives
instead line objectives. Furthermore, the bug in the algorithm which is not incorporating the
avoidance sector should be corrected for more accurate results. This study could serve a
starting point for further validation of the use of RapidPlan for large volume lung cancer
patients with the aim to be implemented in the clinical practice.
50
References
1. Ferlay J, Soerjomataram I, Dikshit R, et al. Cancer incidence and mortality worldwide: Sources, methods and major patterns in GLOBOCAN 2012. doi:10.1002/ijc.29210.
2. Eberhardt WEE, De Ruysscher D, Weder W, et al. 2nd ESMO Consensus Conference in Lung Cancer: locally advanced stage III non-small-cell lung cancer. Ann Oncol. 2015;26(8):1573-1588. doi:10.1093/annonc/mdv187.
3. Wijsman R, Dankers F, Troost EGC, et al. Comparison of toxicity and outcome in advanced stage non-small cell lung cancer patients treated with intensity-modulated (chemo-)radiotherapy using IMRT or VMAT. Radiother Oncol. 2017;122(2):295-299. doi:10.1016/j.radonc.2016.11.015.
4. Palma DA, Senan S, Tsujino K, et al. Predicting Radiation Pneumonitis After Chemoradiation Therapy for Lung Cancer: An International Individual Patient Data Meta-analysis. Int J Radiat Oncol. 2013;85(2):444-450. doi:10.1016/j.ijrobp.2012.04.043.
5. Khalil AA, Hoffmann L, Moeller DS, Farr KP, Knap MM. New dose constraint reduces radiation-induced fatal pneumonitis in locally advanced non-small cell lung cancer patients treated with intensity-modulated radiotherapy. Acta Oncol (Madr). 2015;54(9):1343-1349. doi:10.3109/0284186X.2015.1061216.
6. Pinnix CC, Smith GL, Milgrom S, et al. Predictors of radiation pneumonitis in patients receiving intensity modulated radiation therapy for Hodgkin and non-hodgkin lymphoma. Int J Radiat Oncol Biol Phys. 2015. doi:10.1016/j.ijrobp.2015.02.010.
7. Palma DA, Senan S, Oberije C, et al. Predicting Esophagitis After Chemoradiation Therapy for Non-Small Cell Lung Cancer: An Individual Patient Data Meta-Analysis. Int J Radiat Oncol. 2013;87(4):690-696. doi:10.1016/j.ijrobp.2013.07.029.
8. Mittal BB, Purdy JA, Ang KK. Advances in Radiation Therapy. 1st ed. (Mittal BB, Purdy JA, Ang KK, eds.). Springer US; 1998. doi:10.1007/978-1-4615-5769-2.
9. Bentzen SM. Radiation therapy: intensity modulated, image guided, biologically optimized and evidence based. Radiother Oncol. 2005;77(3):227-230. doi:10.1016/j.radonc.2005.11.001.
10. Murshed H, Liu HH, Liao Z, et al. Dose and volume reduction for normal lung using intensity-modulated radiotherapy for advanced-stage non–small-cell lung cancer. Int J Radiat Oncol. 2004;58(4):1258-1267. doi:10.1016/j.ijrobp.2003.09.086.
11. Liu HH, Wang X, Dong L, et al. Feasibility of sparing lung and other thoracic structures with intensity-modulated radiotherapy for non-small-cell lung cancer. Int J Radiat Oncol Biol Phys. 2004;58(4):1268-1279. doi:10.1016/j.ijrobp.2003.09.085.
12. Otto K. Volumetric modulated arc therapy: IMRT in a single gantry arc. Med Phys. 2007;35(1):310-317. doi:10.1118/1.2818738.
13. Li Y, Wang J, Tan L, et al. Dosimetric comparison between IMRT and VMAT in irradiation for peripheral and central lung cancer. Oncol Lett. 2018;15(3):3735-3745. doi:10.3892/ol.2018.7732.
14. Webb S. Optimisation of conformal radiotherapy dose distribution by simulated annealing. Phys Med Biol. 1989;34(10):1349-1370. doi:10.1088/0031-9155/34/10/002.
51
15. Yuan L, Ge Y, Lee WR, Yin FF, Kirkpatrick JP, Wu QJ. Quantitative analysis of the factors which affect the interpatient organ-at-risk dose sparing variation in IMRT plans. Med Phys. 2012;39(11):6868. doi:10.1118/1.4757927.
16. Das IJ, Cheng C-W, Chopra KL, Mitra RK, Srivastava SP, Glatstein E. Intensity-Modulated Radiation Therapy Dose Prescription, Recording, and Delivery: Patterns of Variability Among Institutions and Treatment Planning Systems. JNCI J Natl Cancer Inst. 2008;100(5):300-307. doi:10.1093/jnci/djn020.
17. Everitt S, Kron T, Fimmell N, et al. Interplanner variability in carrying out three-dimensional conformal radiation therapy for non-small-cell lung cancer. J Med Imaging Radiat Oncol. 2008;52(3):293-296. doi:10.1111/j.1440-1673.2008.01957.x.
18. Nelms BE, Robinson G, Markham J, et al. Variation in external beam treatment plan quality: An inter-institutional study of planners and planning systems. Pract Radiat Oncol. 2012;2(4):296-305. doi:10.1016/j.prro.2011.11.012.
19. Appenzoller LM, Michalski JM, Thorstad WL, Mutic S, Moore KL. Predicting dose-volume histograms for organs-at-risk in IMRT planning. Med Phys. 2012;39(12):7446-7461. doi:10.1118/1.4761864.
20. Fogliata A, Wang P-M, Belosi F, et al. Assessment of a model based optimization engine for volumetric modulated arc therapy for patients with advanced hepatocellular cancer. Radiat Oncol. 2014;9(1):236. doi:10.1186/s13014-014-0236-0.
21. Tol JP, Delaney AR, Dahele M, Slotman BJ, Verbakel WFAR. 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.
22. Delaney AR, Dahele M, Tol JP, Slotman BJ, Verbakel WFAR. Knowledge-based planning for stereotactic radiotherapy of peripheral early-stage lung cancer. Acta Oncol. 2017;56(3):490-495. doi:10.1080/0284186X.2016.1273544.
23. Chanyavanich V, Das SK, Lee WR, Lo JY. Knowledge-based IMRT treatment planning for prostate cancer. Med Phys. 2011;38(5):2515-2522. doi:10.1118/1.3574874.
24. Good D, Lo J, Lee WR, Wu QJ, Yin F-F, Das SK. A Knowledge-Based Approach to Improving and Homogenizing Intensity Modulated Radiation Therapy Planning Quality Among Treatment Centers: An Example Application to Prostate Cancer Planning. Int J Radiat Oncol. 2013;87(1):176-181. doi:10.1016/j.ijrobp.2013.03.015.
25. Moore KL, Brame RS, Low DA, Mutic S. Experience-Based Quality Control of Clinical Intensity-Modulated Radiotherapy Planning. Int J Radiat Oncol. 2011;81(2):545-551. doi:10.1016/j.ijrobp.2010.11.030.
26. Fogliata A, Belosi F, Clivio A, et al. On the pre-clinical validation of a commercial model-based optimisation engine: Application to volumetric modulated arc therapy for patients with lung or prostate cancer. Radiother Oncol. 2014;113(3):385-391. doi:10.1016/j.radonc.2014.11.009.
27. Hussein M, South CP, Barry MA, et al. Clinical validation and benchmarking of knowledge-based IMRT and VMAT treatment planning in pelvic anatomy. Radiother Oncol. 2016;120(3):473-479. doi:10.1016/j.radonc.2016.06.022.
28. Nwankwo O, Mekdash H, Sihono DSK, Wenz F, Glatting G. Knowledge-based radiation therapy (KBRT) treatment planning versus planning by experts: validation of a KBRT algorithm
52
for prostate cancer treatment planning. Radiat Oncol. 2015;10(1):111. doi:10.1186/s13014-015-0416-6.
29. Chin Snyder K, Kim J, Reding A, et al. Development and evaluation of a clinical model for lung cancer patients using stereotactic body radiotherapy (SBRT) within a knowledge-based algorithm for treatment planning. J Appl Clin Med Phys. 2016;17(6):263-275. doi:10.1120/jacmp.v17i6.6429.
30. Fogliata A, Nicolini G, Clivio A, et al. 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;10:220. doi:10.1186/s13014-015-0530-5.
31. Fogliata A, Nicolini G, Bourgier C, et al. Performance of a Knowledge-Based Model for Optimization of Volumetric Modulated Arc Therapy Plans for Single and Bilateral Breast Irradiation. Gilhuijs KGA, ed. PLoS One. 2015;10(12):e0145137. doi:10.1371/journal.pone.0145137.
32. Verbakel WFAR, Van Reij E, Ladenius-Lischer I, Cuijpers JP, Slotman BJ, Senan S. Clinical application of a novel hybrid intensity-modulated radiotherapy technique for stage III lung cancer and dosimetric comparison with four other techniques. Int J Radiat Oncol Biol Phys. 2012;83(2):e297-e303. doi:10.1016/j.ijrobp.2011.12.059.
33. Chan OSH, Lee MCH, Hung AWM, Chang ATY, Yeung RMW, Lee AWM. The superiority of hybrid-volumetric arc therapy (VMAT) technique over double arcs VMAT and 3D-conformal technique in the treatment of locally advanced non-small cell lung cancer - A planning study. Radiother Oncol. 2011;101(2):298-302. doi:10.1016/j.radonc.2011.08.015.
34. Silva SR, Surucu M, Steber J, Harkenrider MM, Choi M. Clinical Application of a Hybrid RapidArc Radiotherapy Technique for Locally Advanced Lung Cancer. Technol Cancer Res Treat. 2016. doi:10.1177/1533034616670273.
35. Tol JP, Dahele M, Delaney AR, Slotman BJ, Verbakel WFAR. Can knowledge-based DVH predictions be used for automated, individualized quality assurance of radiotherapy treatment plans? Radiat Oncol. 2015;10(1):234. doi:10.1186/s13014-015-0542-1.
36. Delaney AR, Dahele M, Tol JP, Kuijper IT, Slotman BJ, Verbakel WFAR. Using a knowledge-based planning solution to select patients for proton therapy. Radiother Oncol. 2017;124(2):263-270. doi:10.1016/j.radonc.2017.03.020.
37. Blom GJ, Verbakel WF a R, Dahele M, Hoffmans D, Slotman BJ, Senan S. Improving radiotherapy planning for large volume lung cancer: a dosimetric comparison between hybrid-IMRT and RapidArc. Acta Oncol. 2015;54(3):427-432. doi:10.3109/0284186X.2014.963888.
38. Tol JP, Dahele M, Doornaert P, Slotman BJ, Verbakel WFAR. Toward optimal organ at risk sparing in complex volumetric modulated arc therapy: An exponential trade-off with target volume dose homogeneity. Med Phys. 2014;41(2):021722. doi:10.1118/1.4862521.
39. Wolff D, Stieler F, Welzel G, et al. Volumetric modulated arc therapy (VMAT) vs. serial tomotherapy, step-and-shoot IMRT and 3D-conformal RT for treatment of prostate cancer. Radiother Oncol. 2009;93(2):226-233. doi:10.1016/j.radonc.2009.08.011.
40. No Title. https://varian.force.com/CpWebSummary?Id=a0OE000000NrOvaMAF.
41. Varian Medical Systems. Eclipse Photon and Electron Algorithms Reference Guide Version
53
15.1. Palo Alto, CA 94304-1038 United States of America; 2016. doi:P1015026-002-B.
42. Delaney AR, Tol JP, Dahele M, Cuijpers J, Slotman BJ, Verbakel WFAR. Effect of Dosimetric Outliers on the Performance of a Commercial Knowledge-Based Planning Solution. Int J Radiat Oncol Biol Phys. 2016;94(3):469-477. doi:10.1016/j.ijrobp.2015.11.011.
43. Varian Medical Systems. Eclipse Photon and Electron Reference Guide Version 13.5. Palo Alto, CA 94304-1038 United States of America; 2014. doi:P1005653-002.
54
APPENDIX A
Geometry Expected Dose calculation algorithm not incorporating the avoidance sector.
To investigate the reason of full-arc plans in the h-RA model have higher GED-PC1, the
GED-PC1 for CL was ploted against the following geometrical features: overlap with target,
in-field volume, target volume and mean dose (Figure A1). Full-arc with avoidance sector
plans (red squares) deviated from the rest of the plans, despite not having different
geometrical features.
Figure A1: Overlap with target, in-field volume, target volume and mean dose of CL plotted
against GED-PC1 for h-RA model. Red squares include the full-arc with avoidance sector
plans.
To further examine the handling of avoidance sectors by the model, for two patients
three treatment plans were created and interactively optimized. The first treatment plan
incorporated a full-arc and avoidance sector (a), the second a partial-arc covering the same
irradiative angles of the first plan (b) and the third using full irradiating arc without any
avoidance sector. (c).These plans were subsequently added to the f-RA model, and the model
was re-trained to visualize where these plans lie in the generated regression plot for the CL
(Figure A2). The full-arc without avoidance sector has higher GED-PCS compared to the
other two plans, as expected. One would expect that when the irradiating field directions are
55
the same for both plans, the GED and consequently the GED-PCS1 would be the same for the
two plans. However, after model training, the avoidance sector plan had a considerably
higher GED-PCS1, suggesting that the GED is not accurately calculated. The RapidPlan
provided geometric plots (Figure A2 (B) and (C)) show that the geometrical features (CL
volume, overlap with target, out-of-field volume and target volume) are the same for the three
plans. The only other factor that could cause the GED to be different is the field set-up.
Figure A2: (A) Regression plot of CL of the f-RA model. (B) Geomteric plots for patient 1,
for the (a) partial-arc plan, (b) full-arc with avoidance sector plan (c) full-arc plan. (C)
Geometric plots for patient 2 for the corresponding plans.
56
This bug in the algorithm, had an effect on the generation of predictions. The DVH
estimation algorithm does not take into account the avoidance sector in the optimization
window, and thus the predictions for the CL and TL-PTV are high (1st prediction in Figure
A3). After optimizing the plan incorporating the avoidance sector, and then using the
optimized plan to generate new predictions, the predictions appear to be lower (2nd prediction
in Figure A3). The 1st prediction corresponds to the case c/full-arc without avoidance sector
in Figure A2, while the 2nd prediction corresponds to the case b/full-arc with avoidance
sector. If we use partial-arcs, generate the predictions and optimize the plan, the predictions
seem to be far lower than the achieved DVH.All the f-RA MBPs in the study were optimized
two times to incorporate the avoidance sector. However, since when using partial-arc the
predictions are even lower, the problem is not completely solved.
The problem was communicated to a Varian representative who gave the explanation
to our observations. The GED calculation algorithm does not recognize the avoidance sectors
and ‘discretize’ the arc fields. If the difference in gantry angle is less than 5 degree between
two control points -beam directions-, the control points are used for the discretization, but if
the difference is more than 5 degree, additional control points are interpolated between the
two control points. The new control points are equally spaced and the amount is deduced so
that the new spacing is less or equal than 5 degree. When we start with new fields to generate
predictions for a plan (1st prediction), the control points are spaced in 5 degree along the full
arc, so the predictions are generated as if there is no avoidance sector. When optimizing the
plan, we manually set the control points in around 2 degree interval in the irradiation-arc and
no control points in the avoidance sector. When using the optimized plan -whose fields have
already control points- to generate predictions (2nd prediction), any avoidance sector is
considered as a long jump between two control points and additional control points are
generated with 5 degree spacing. Therefore, the 2nd prediction is generated as if in the
avoidance sector the ‘density’ of control points has been dropped from one control point in
every 2 degree into one control point in every 5 deg. This is actually affecting the GED since
it assumes less dose from those directions. If we use partial arcs, there are no control
points/beamlets at all in the avoidance sector directions. The correct functionality would be
that the partial-arc fields and the full-arc fields with avoidance sector should give the same
results.
However, the problem is not encountered only in the predictions, it is also in the
models since the same GED calculation algorithm is used for the model training. Using
57
partial-arc instead of avoidance sector in the MBPs does not solve the problem because our
models is mainly consisted of avoidance sector plans.
Figure A3: DVH predictions for a new plan (1st prediction) and for an already optimized
plan (2nd prediction) of the same patient, with full-arc with avoidance sector, for CL and TL-
PTV.
58
APPENDIX B
Supplementary material
Figure B1: Predicted DVH-ranges (shaded areas), achieved MBP (black) and manual plan
DVH-lines (red) of Contralateral Lung with the f-RA method for the 10 test patients.
59
Figure B2: Predicted DVH-ranges (shaded areas), achieved MBP (black) and manual plan
DVH-lines (red) of TL-PTV with the f-RA method for the 10 test patients.