Key words: dose optimization, IMRT, inverse planning, treatment planning, VMAT
389 Med. Phys. 44 (2), February 2017 0094-2405/2017/44(2)/389/8 2016 American Association of Physicists in Medicine 389
390 Wang et al.: Autonomous treatment planning 390
employed to train the system to provide the best model with a commercial treatment planning system (TPS) to imple-
parameters. The parameters are then used as input of subse- ment the proposed strategy that mimics a planners interac-
quent inverse planning. The method has recently been applied tive planning and decision-making process in searching for a
to predict the weighting factors911 and the prescription sensible solution. The approach here enables us to leverage
DVHs1215 needed for driving an inverse planning calcula- the sophisticated software subroutines (e.g., dose calculation,
tion. We emphasize that all these approaches only use prior optimization, image registration, etc.) existing already in a
knowledge-derived parameters (i.e., either weighting or pre- clinical-grade TPS to test the new algorithm. By implement-
scription or both) to warm start the inverse planning, ing the outer-loop optimization strategy in this programming
instead of using them to guide the plan search throughout the environment, we demonstrate that IMRT/VMAT treatment
optimization process. Additionally, the predicted DVHs pre- planning can be readily automated.
scription by machine learning may not always be physically
realizable. For completeness, we mention that heuristic opti-
2. METHODS AND MATERIALS
mization14,1621 and multiobjective optimization2224 have
also been developed to facilitate inverse planning process and Our treatment planning system consisted of following
these algorithms can, in principle, be improved with incorpo- components: (i) a clinical TPS; (ii) a robust formulation of
ration of prior knowledge, especially population-based data. plan evaluator or decision-function; (iii) an outer-loop opti-
The purpose of this work is to develop an autonomous mization independent of the clinical TPS to assess a TPS-
treatment planning technique with effective use of a spectrum generated plan and to drive the search toward a solution con-
of prior plans that share similar anatomical features with the sistent with the decision-function; and (iv) a programming
case under planning. Instead of using a simplified research method to query and interact with the TPS using Microsoft
inverse planning platform,2528 which often ignores some (MS) Visual Studio Coded UI, which is applied to record
important geometric and physical factors, we proceed with an some common planner-TPS interactions as subroutines. The
API-like programming environment capable of interacting following summarizes the details of the technique.
FIG. 1. Flowchart of the autopiloted plan optimization scheme. An outer-loop decision-function analyzes the TPS plan and feeds the TPS optimizer with updated
parameters for iterative improvement of the treatment plan. The algorithm terminates if either (i) ck = 0, or (ii) no improvement has been made in consecutive
iterations. [Color figure can be viewed at wileyonlinelibrary.com]
planning process. However, the final solution does not rely considering the pilot nature of the study, the number is
on the selection of the initial selection of starting plan. Dur- reasonable.
ing the calculation, instead of letting the DVH segment value
Crj stop anywhere when its value is inside the range defined
by Crjrefl refh
and Crj , we continuously push the Crj ref
value 3. RESULTS
toward a lower value in the predefined range by changing the 3.A. Five-eld IMRT prostate treatment
reference value.
In Fig. 2, we show the band of DVH representing the pre-
ferred range and distribution of the resultant DVH curve for
2.C. Evaluation the prostate case under planning. To illustrate the progressive
The above technique is applied to plan two clinical improvement of the autopilot process, in Fig. 3 we show the
cases: a five-field IMRT prostate case and a VMAT DVH results of involved structures for the 1st, 7th, and 14th
head-and-neck (HN) case. In the prostate IMRT case, iteration. The improvement saturate after about 14 iterations
6 MV photon energy is used and the beam angles are and the calculation thus terminates after the 14th iteration. At
0o, 50o, 100o, 260o, and 310o, respectively. 78 Gy is pre- each outer-loop iteration, the TPS parameters are adjusted,
scribed to cover V95 of the PTV in 39 fractions. For which leads to a different dose distribution. While these plans
comparison, the resultant dose distributions of the autopi- are on the Pareto front, the autoplanning process here
lot scheme are compared with the corresponding plans explores plans beyond the traditional Pareto front. For exam-
used for clinical treatments. To construction of the refer- ple, when a hot spot presents in PTV, a tuning structure
ence plan library, 15 previously treated prostate cancer would be created automatically by converting the correspond-
patients are selected using the procedure outlined in Sec- ing isodose curve into a structure and this process would
tion 2.B. For the HN case, two 360o 6 MV VMAT arcs drive the plan away from the traditional Pareto surface, lead-
are used. Nine previously treated HN cases are selected ing to improvements in all DVHs. Figure 4 shows the DVH
as reference plans. We acknowledge that the database can comparison between the clinical and autopiloted plans for the
be enlarged to better cover the variations in patient anat- case. Figure 5 shows the isodose distribution of the two
omy. Given that the cases selected to form the library plans. Only minor discrepancy is seen between the autopi-
are reasonably close to the cases under planning and loted plan obtained under the guidance of the population-
FIG. 2. Plots of reference DVHs of bladder (a), rectum (b), and PTV (c) for the prostate case under planning. The ensemble of DVHs represents the preferred
range of the resultant DVH curve.
FIG. 3. DVHs of the bladder, rectum, and prostate at iteration #1 (dotted), #7 (dashed), and #14 (solid). A systematic improvement in the DVHs is observed.
based library of reference plans and clinical plan generated discrepancy is seen between the autopiloted plan obtained
by a human planner independently. under the guidance of the population-based library of refer-
ence plans and clinical plan generated by a human planner
independently.
3.B. HN VMAT case
Computationally, depending on the complexity of the
In Fig. 6, we show the ensemble of DVHs of the spinal case, it takes about a few hours to complete a plan selection
cord, brainstem and PTV of the reference plans obtained process, but this can be improved with improved program-
using the method of Section 2.B for the HN case under plan- ming, and, in the future, better integration with the commer-
ning. Figure 7 shows the DVH comparison between the clini- cial TPS. However, comparing with the current manual
cal and autopiloted planning for the case, and Fig. 8 shows planning process, which could take days of a dosimetrists
the isodose distributions of the two plans. Again, only minor time for clinically challenging cases, the proposed method
FIG. 4. A comparison of DVHs of the clinical and autopiloted plans for the five-field IMRT prostate case. The dashed and solid curves represent the DVHs of
clinical and autopiloted plans, respectively.
FIG. 5. Side-by-side comparison of the isodose distributions of autopiloted (right) and clinical (left) plans for the prostate case. [Color figure can be viewed at
wileyonlinelibrary.com]
FIG. 6. Plots of reference DVH curves of spinal cord (a), brainstem (b), and PTV (c) for the head-and-neck case under planning. The ensemble of DVHs presents
the preferred range of the resultant DVH curve.
eliminates the manual trial-and-error modification of the opti- multiple model parameters.1,10,27,3235 In the past two dec-
mization parameters and facilitates the clinical workflow. ades, much of the efforts in research and commercial product
development have been focused on finding a better function
with mathematical constraints to provide clinically sensible
4. DISCUSSION
solution. While the formulation of objective function and the
In inverse planning, an optimized solution is obtained technique used to search through the solution space under the
under the guidance of an objective function containing guidance of the objective function are important to the
FIG. 7. A comparison of DVHs of the clinical and autopiloted plans for the two-arc head and neck VMAT case. The dashed and solid curves represent the DVHs
of clinical and autopiloted plans, respectively.
FIG. 8. Side-by-side comparison of the isodose distributions of autopiloted (right) and clinical (left) plans for the head-and-neck case. [Color figure can be
viewed at wileyonlinelibrary.com]
success of inverse planning, clinical IMRT/VMAT planning inverse planning process. The research here is directly
has been handicapped by the involvement of multiple model translatable to clinical practice as the backbone of the plan
parameters that necessitates trial-and-error determination. In optimization is built upon a commercial TPS. Additionally,
this paper, we proposed an effective technique to autopilot the approach is quite general and allows us to incorporate
the VMAT/IMRT planning process in a clinical TPS plat- empirical judgment and population-based prior knowledge
form. The automation is realized by combining the function- into the plan selection process. The utility of the approach
alities of the black-box TPS and a decision-function has been demonstrated successfully by using two clinical
incorporating population-based prior knowledge. The calcu- cases. With the increased interest in using prior knowledge
lation is analogous to the planning process of a human plan- in radiation oncology applications, the data analytics and
ner, with a candidate plan assessed iteratively by the decision-making method may prove to be useful to facili-
decision-function each time after the Eclipse optimization is tate clinical workflow.
done.1,2,10 During the autopiloted planning, the C# program
interacts with TPS continuously to extract the updated infor-
ACKNOWLEDGMENTS
mation and to instruct on what to do next until a satisfactory
plan is obtained. This work is partially supported by NIH (5R01
In addition to the effective utilization of population-based CA176553) and Varian Medical Systems.
prior data, a technical innovation of the work here is the use
of Coded UI to recording the mouse clicks/keystrokes as exe-
CONFLICT OF INTEREST
cutable subroutines for specific tasks during planning. By
doing so, we can extract the updated information of treatment Lei Xing is the principal investigator of a master research
planning and interact with the TPS effectively to autopilot the agreement (MRA) with Varian Medical Systems.
process. In practice, because the Windows platform encom-
passes a large number of TPS and other clinical application
a)
software, the presented method should be valuable for future Author to whom correspondence should be addressed. Electronic mail:
lei@stanford.edu.
clinical research and practice. We note that, while Coded UI
is designed for the Windows environment, the principles and
strategies proposed here are quite broad. There are analogous REFERENCES
software tools for other operating systems. For Unix and
1. Xing L, Li JG, Donaldson S, Le QT, Boyer AL. Optimization of impor-
Linux environment, other record-and-playback services such tance factors in inverse planning. Phys Med Biol. 1999;44:25252536.
as UNIX Session Recorder (Sikuli, Linux Desktop Testing 2. Xing L, Li JG, Pugachev A, Le QT, Boyer AL. Estimation theory and
Project) can be employed.36 model parameter selection for therapeutic treatment plan optimization.
Med Phys. 1999;26:23482358.
The programming environment described here is highly 3. Yu Y, Zhang JB, Cheng G, Schell MC, Okunieff P. Multi-objective opti-
interactive, making it easy to principle-testing and prototyp- mization in radiotherapy: applications to stereotactic radiosurgery and
ing. Currently, the development of treatment planning algo- prostate brachytherapy. Artif Intell Med. 2000;19:3951.
rithm(s) is often done in simplified software platforms2528 4. Zarepisheh M, Long T, Li N et al. A DVH-guided IMRT optimization
algorithm for automatic treatment planning and adaptive radiotherapy
without considering some important geometric and physical replanning. Med Phys. 2014;41:061711.
factors. The approach here enables researchers to leverage the 5. Shepard DM, Earl MA, Li XA, Naqvi S, Yu C. Direct aperture opti-
sophisticated software subroutines existing already in a clini- mization: a turnkey solution for step-and-shoot IMRT. Med Phys.
cal-grade TPS to test new ideas without repeating some well- 2002;29:10071018.
6. Earl MA, Shepard DM, Naqvi S, Li XA, Yu CX. Inverse planning for
known tasks. The approach may also facilitate the translation intensity-modulated arc therapy using direct aperture optimization. Phys
of research to clinical practice. Finally, we note that the Med Biol. 2003;48:10751089.
implementation here represents only one of many possible 7. Chanyavanich V, Das SK, Lee WR, Lo JY. Knowledge-based IMRT
applications of the proposed strategy. The approach is appli- treatment planning for prostate cancer. Med Phys. 2011;38:25152522.
8. Schreibmann E, Xing L. Feasibility study of beam orientation class-
cable to streamline other clinical tasks in TPS or other soft- solutions for prostate IMRT. Med Phys. 2004;31:28632870.
ware tools. We also note that practical issues such as error 9. Good D, Lo J, Lee WR, Wu QJ, Yin FF, Das SK. A knowledge-based
handling, such as unexpected windows popup during the approach to improving and homogenizing intensity modulated radiation
therapy planning quality among treatment centers: an example applica-
autopiloted planning process, are not discussed. In principle,
tion to prostate cancer planning. Int J Radiat Oncol Biol Phys.
a preventive action can be taken by adding some checking 2013;87:176181.
points in the C# program of autopiloted planning to monitor 10. Lee T, Hammad M, Chan TC, Craig T, Sharpe MB. Predicting objective
the potential popup window(s). function weights from patient anatomy in prostate IMRT treatment plan-
ning. Med Phys. 2013;40:121706.
11. Boutilier JJ, Lee T, Craig T, Sharpe MB, Chan TC. Models for predict-
CONCLUSION ing objective function weights in prostate cancer IMRT. Med Phys.
2015;42:15861595.
Inverse planning in modern radiation therapy involves 12. Zarepisheh M, Li R, Ye Y, Xing L. Simultaneous beam sampling and
aperture shape optimization for SPORT. Med Phys. 2015;42:10121022.
multiple steps of manual operation and is known to be a 13. Appenzoller LM, Michalski JM, Thorstad WL, Mutic S, Moore KL.
time-consuming process. In this work, we propose an Predicting dose-volume histograms for organs-at-risk in IMRT planning.
autonomous treatment planning strategy to facilitate the Med Phys. 2012;39:74467461.
14. Lian J, Yuan L, Ge Y et al. Modeling the dosimetry of organ-at-risk in 25. Deasy JO, Blanco AI, Clark VH. CERR: a computational environment
head and neck IMRT planning: an intertechnique and interinstitutional for radiotherapy research. Med Phys. 2003;30:979985.
study. Med Phys. 2013;40:121704. 26. Tewell MA, Adams R. The PLUNC 3D treatment planning system: a
15. Wu Q, Yuan L, Li T, Ying F, Ge Y. Knowledge-based organ-at-risk spar- dynamic alternative to commercially available systems. Med Dosim.
ing models in IMRT planning. Pract Radiat Oncol. 2013;3:S1S2. 2004;29:134138.
16. Purdie TG, Dinniwell RE, Letourneau D, Hill C, Sharpe MB. Auto- 27. Kim H, Li R, Lee R, Xing L. Beams-eye-view dosimetrics (BEVD)
mated planning of tangential breast intensity-modulated radiotherapy guided rotational station parameter optimized radiation therapy (SPORT)
using heuristic optimization. Int J Radiat Oncol Biol Phys. 2011;81:575 planning based on reweighted total-variation minimization. Phys Med
583. Biol. 2015;60:N71N82.
17. Purdie TG, Dinniwell RE, Fyles A, Sharpe MB. Automation and inten- 28. Dong P, Ungun B, Boyd S, Xing L. Optimization of rotational arc sta-
sity modulated radiation therapy for individualized high-quality tangent tion parameter optimized radiation therapy. Med Phys. 2016;43:4973.
breast treatment plans. Int J Radiat Oncol Biol Phys. 2014;90:688695. 29. Johnson B. Professional Visual Studio. Indianapolis, IN: John Wiley &
18. Liu H, Wu Q. Evaluations of an adaptive planning technique incorporat- Sons; 2015.
ing dose feedback in image-guided radiotherapy of prostate cancer. Med 30. https://en.wikipedia.org/wiki/Microsoft_UI_Automation#cite_note-2,
Phys. 2011;38:63626370. 2016.
19. Shiraishi S, Tan J, Olsen LA, Moore KL. Knowledge-based prediction 31. Schreibmann E, Xing L. Narrow band deformable registration of prostate
of plan quality metrics in intracranial stereotactic radiosurgery. Med magnetic resonance imaging, magnetic resonance spectroscopic imag-
Phys. 2015;42:908. ing, and computed tomography studies. Int J Radiat Oncol Biol Phys.
20. Amit G, Purdie TG, Levinshtein A et al. Automatic learning-based 2005;62:595605.
beam angle selection for thoracic IMRT. Med Phys. 2015;42:19922005. 32. Li RP, Yin FF. Optimization of inverse treatment planning using a fuzzy
21. Schreibmann E, Fox T. Prior-knowledge treatment planning for volumet- weight function. Med Phys. 2000;27:691700.
ric arc therapy using feature-based database mining. J Appl Clin Med 33. Wu Q, Djajaputra D, Wu Y, Zhou J, Liu HH, Mohan R. Intensity-modu-
Phys. 2014;15:4596. lated radiotherapy optimization with gEUD-guided dose-volume objec-
22. Cotrutz C, Lahanas M, Kappas C, Baltas D. A multiobjective gradient- tives. Phys Med Biol. 2003;48:279291.
based dose optimization algorithm for external beam conformal radio- 34. Chan TC, Bortfeld T, Tsitsiklis JN. A robust approach to IMRT opti-
therapy. Phys Med Biol. 2001;46:21612175. mization. Phys Med Biol. 2006;51:25672583.
23. Kamran SC, Mueller BS, Paetzold P et al. Multi-criteria optimization 35. Zhang P, Yorke E, Hu YC, Mageras G, Rimner A, Deasy JO. Predictive
achieves superior normal tissue sparing in a planning study of intensity- treatment management: incorporating a predictive tumor response model
modulated radiation therapy for RTOG 1308-eligible non-small cell lung into robust prospective treatment planning for non-small cell lung can-
cancer patients. Radiother Oncol. 2016;118:515520. cer. Int J Radiat Oncol Biol Phys. 2014;88:446452.
24. Unkelbach J, Bortfeld T, Craft D et al. Optimization approaches to volu- 36. https://www.ibm.com/developerworks/community/blogs/nix/entry/the_
metric modulated arc therapy planning. Med Phys. 2015;42:13671377. unix_script_command_a_command_recorder2?lang=en, 2016.