VU Research Portal

(1)

Knowledge-based treatment planning for radiotherapy

Delaney, A.R.

2019

document version

Publisher's PDF, also known as Version of record

Link to publication in VU Research Portal

citation for published version (APA)

Delaney, A. R. (2019). Knowledge-based treatment planning for radiotherapy.

General rights

Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain

• You may freely distribute the URL identifying the publication in the public portal ?

Take down policy

If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.

E-mail address:

vuresearchportal.ub@vu.nl

(2)

KNOWLEDGE - BASED

TREA

TMENT

PLANNING FOR RADIOTHERAPY

(3)

(4)

534143-L-sub01-bw-Delaney 534143-L-sub01-bw-Delaney 534143-L-sub01-bw-Delaney 534143-L-sub01-bw-Delaney Processed on: 7-8-2019 Processed on: 7-8-2019 Processed on: 7-8-2019

Processed on: 7-8-2019 PDF page: 1PDF page: 1PDF page: 1PDF page: 1

Knowledge-based

treatment planning

for radiotherapy

(5)

Printed by: Ipskamp Printing ISBN: 978-94-028-1655-6

Lay-out: Proefschrift-AIO, Talitha Vlastuin © Copyright: Alexander Richard Delaney, 2019

(6)

VRIJE UNIVERSITEIT

Knowledge-based

treatment planning

for radiotherapy

ACADEMISCH PROEFSCHRIFT

ter verkrijging van de graad Doctor of Philosophy aan de Vrije Universiteit Amsterdam,

op gezag van de rector magnificus prof.dr. V. Subramaniam, in het openbaar te verdedigen ten overstaan van de promotiecommissie

van de Faculteit der Geneeskunde op donderdag 26 september 2019 om 9.45 uur

in de aula van de universiteit, De Boelelaan 1105

door

Alexander Richard Delaney

(7)

(8)

Introduction

Evaluation of a knowledge-based planning solution for head and neck cancer

Effect of Dosimetric Outliers on the Performance of a Commercial Knowledge-Based Planning Solution

Knowledge-based planning for stereotactic radiotherapy of peripheral early-stage lung cancer

Is accurate contouring of salivary and swallowing structures necessary to spare them in head and neck VMAT plans? Using a knowledge-based planning solution to select patients for proton therapy

Evaluation of an automated proton planning solution Automated Knowledge-Based Intensity-Modulated Proton Planning: An International Multicenter Benchmarking Study Discussion and Future Directions

(9)

(10)

7

1

(11)

(12)

9

1

Whilst infectious diseases were among the predominant causes of death in the

19th_{century [1], the 20}th_{century saw the continuous growth in death from}

non-communicable illnesses such as heart disease and cancer. Cancer, as a single entity, outranked deaths from all other major causes of death in 2011 [2]. Furthermore, the global increase in cancer is not expected to abate with projected numbers of new cases to surpass 20 million by 2025, compared with an estimated 14.1 million in 2014 [2].

This thesis focuses on radiotherapy treatment planning for head and neck cancer (HNC), and to a lesser extent lung cancer. Stratifying cancer types according to rate of incidence, HNC accounted for an estimated 140,000 new cases and 63,500 deaths in Europe in 2012 and equates to about 4% of all cancers arising in this part of the world [3]. The predominant causes of HNC include behavioral risk factors such as tobacco/alcohol use [4], and an increasing proportion related to the human papillomavirus (HPV). The prevalence of HPV in oropharyngeal cancer patients was around 40% and 72% for studies recruiting prior to 2000 and after 2005, respectively [5]. Lung cancer is the most frequently occurring and one of the most lethal, with > 1.8 million new cases and 1.6 million deaths globally in 2012. In Europe, lung cancer accounted for approximately 410,000 new cases and 291,000 deaths in 2012, equating to about 12% of all cancers [2]. Lung cancer can be classified histologically as either small cell lung cancer or non-small cell lung cancer (NSCLC) with the latter accounting for about 85% of cases [6]. The predominant cause of lung cancer is cigarette smoking, with the number of “pack-years” (product of amount smoked and duration of smoking) of importance [2]. Based on recent data, the 5-year overall survival rate of HNC (oro/naso/hypo-pharynx) in the Netherlands is 45%, whereas it is only 19% for lung cancer [7]. However, estimates for 5-year survival of stage 1 non-small cell lung cancer range from 68% to 92% [8].

(13)

10

and fractionation schedule; a computed tomography (CT) scan of the patient in the intended treatment position is obtained; the position and visual extent of the gross tumor (gross tumor volume, GTV) is delineated on this CT scan with the help of additional clinical and imaging information; the GTV is then expanded to account for the spread of sub-clinical disease, creating the clinical target volume (CTV). Finally, to account for any uncertainties in treatment planning/delivery the planning target volume (PTV) is delineated.

To reach targets in the body, a certain amount of radiation will unavoidably be deposited in surrounding healthy organs/tissues (“organs at risk”, OARs). This is caused mostly by the entrance and exit dose of the beam, and dose from radiation which scatters from within the beam to surrounding tissue. This dose to OARs can result in unwanted toxicities which can worsen a patient’s quality of life and in the most serious situations, shorten it. During radiotherapy of HNC, the salivary glands may receive dose which can result in inadequate function and associated dry mouth or “xerostomia”. Similarly, dose to swallowing muscles may lead to swallowing dysfunction or “dysphagia”. For several organs, such toxicities have been attempted to be correlated with dose-volume characteristics that enable Normal Tissue Complication Probability (NTCP) models to be derived - the risk for a given side effect typically increases with increasing dose and increasing volume of a given OAR receiving a certain dose. For instance, at a mean dose of 40Gy, there is a 50% probability of parotid gland flow reduction to <25% of pre-radiotherapy levels [10]. In radiotherapy of peripheral early-stage non-small cell lung cancer, lesions can abut normal tissue such as the thoracic wall and the volume of the thoracic wall receiving above 30 Gy may be predictive of chest wall pain/rib fractures [11]. Improvements in radiotherapy delivery have therefore been predicated on delivering a more conformal and targeted treatment, minimizing irradiation of normal tissues. This is symbolized by the progression of external beam photon radiotherapy from 2D radiotherapy, to 3D conformal radiation therapy (3DCRT) to intensity modulated radiation (IMRT) and volumetric modulated radiation therapy (VMAT).

(14)

11

1

treatment field. IMRT relies heavily on the synergistic effect of multiple beams where an inhomogeneous dose delivered by one beam direction is compensated for by modulated beams from other directions. This leads to achieving an even more conformal high-dose region, at the expense of more low-dose spread outside the target volume [13]. This can provide complex dose distributions which allow steep dose gradients in scenarios where an OAR is adjacent to the target volume. Volumetric modulated arc therapy (VMAT) is an arc variant of IMRT, in which static treatment fields are replaced by the gantry rotating continuously around the patient delivering continuously modulated beam apertures from many more angles [14]. VMAT facilitates rapid delivery of treatment, with dual arc plans from one vendor requiring <3 minutes in contrast to 8-12 minutes for IMRT plans using 7 static fields. Such improvements may reduce the risks of intrafraction movements and improve patient comfort [15].

Figure 1. Left: Typical treatment machine (linear accelerator, www.radiologyinfo.org) Right:

Multi-leaf collimator (www.onko-i.si)

(15)

12

scattered proton therapy, which is limited in its ability to tailor dose distributions, this too remains to be seen [20]. The recently developed intensity modulated proton therapy (IMPT) is currently considered the gold standard for delivering proton radiotherapy. IMPT uses active scanning technology to position Bragg peaks of individually weighted proton beams throughout the target. Similar to IMRT, IMPT relies on the synergistic effects of individual fields which, separately deliver a heterogeneous dose, but combined, lead to a homogeneous dose across the PTV.

Figure 2. Depth-dose curves for photon and proton beams (Leeman JE, et al. Lancet Oncol. 2017

May;18(5):e254-e265.)

(16)

13

1

and the optimization process are subject to considerable variation between clinicians and planners, which can influence the quality of the treatment plan.

Contouring has traditionally been a human-dependent task, with accurate delineation of complex disease sites like the head and neck often time- and labor-intensive. Furthermore, the clarity of the target or an OAR on a CT scan, the experience of a physician and simply the level of attention paid to this process means that manual contours are often prone to variability [22]. A number of automated solutions have been proposed to mitigate such variability and reduce the resources/time associated with manual contouring. Atlas-based segmentation is one such approach, however, this solution still typically requires post-segmentation editing of the automatically generated contours and therefore necessitates the presence of a physician, limiting potential for gains in time and efficiency [23–25]. Contouring may not only be needed before the start of radiotherapy, but sometimes during it as well. Patient, target and OAR geometries can change due to factors such as weight loss or tumor regression. Ideally, such patients would receive adaptive planning, with delineations altered to accommodate transformations in geometry followed by a new treatment plan, however, the obstacles associated with human-driven contouring means that such tailored approaches are not always feasible for a radiotherapy department.

Similar to manual contouring practices, manual treatment planning also exhibits considerable variation. The differing geometrical characteristics of each patient, such as the distance between an OAR and the PTV and OAR-PTV overlap, affect the level of achievable sparing [26]. Due to these patient-specific differences no universal template of dose-volume constraints can be applied to patients if the goal is to provide near-optimal plan quality. If the chosen constraints are too ambitious, this may lead to under-dosage of PTVs, or if they are too lax, it may lead to sub-optimal OAR sparing, and thus subsub-optimal plan quality. Therefore, constraints would ideally be tailored per patient. If this is done manually, then the appropriate choice of constraints is largely dependent on the experience and skill of the planner. As Nelms et al. have pointed out, however, planner skill varies and this contributes to the observed variability in treatment plan quality [27]. Such variation and inconsistencies could have a significant impact on the toxicity that the patient experiences after radiotherapy, and thus impact the treatment outcome. For patients participating in clinical trials, suboptimal plan quality can furthermore detract from the results of such clinical trials. Peters et al. documented the impact of radiotherapy quality in HNC, finding that deviation from protocol or clinical guidelines could sometimes result in substantial adverse effects on tumor control [28].

(17)

14

optimization (MCO) is one example of an approach to more automated treatment planning. Rather than iteratively re-optimizing one treatment plan to move closer to the desired criteria of a physician, MCO uses a database of treatment plans which satisfy different planning objectives, constituting what is often referred to as the “Pareto curve” The plans on this curve are in theory pareto efficient/optimal, in that any change in dose to one OAR results in a tradeoff with another. MCO allows the treatment planner and physician to move between points on this pareto curve, intuitively navigating between tradeoffs, and choosing a plan which best meets the desired criteria [29,30]. However, the introduction of MCO into routine clinical workflow, in an effective manner, may not be effortless. The intention of pareto surface navigation is to hand the tradeoff decision to the decision maker, or the radiation oncologist. Departments practicing traditional forms of treatment planning for many years may take some time in transitioning to a point where the physician takes an active role in treatment planning. This is significant when we consider that a treatment planner may select a plan that meets the prescribed goals, whereas a physician may have navigated to a different part of the pareto surface. A number of MCO solutions for treatment planning have been created including iCycle, from Erasmus medical center in Rotterdam. In iCycle, automated plan generation, using MCO, is guided by a user-defined wish-list comprising objectives and hard constraints [31–33]. Commercial vendors have also introduced MCO into their TPS showing comparable/improved plan quality when compared with manual optimization methodologies [34]. A number of alternative automated solutions have also been devised. Tol et al. developed an automated alternative to manual, interactive optimization (automatic interactive optimizer, AIO). As opposed to interactively adjusting dose-volume objectives to meet specified planning goals, AIO automatically alters the location of objectives throughout the optimization process. AIO has shown at least comparable or improved dosimetry over manually optimized treatment plans in the head and neck [35] and lung [36]. Additional approaches include Autoplanning (Philips Medical Systems), which attempts to iteratively fine tune target coverage and OAR sparing through the creation of additional structures which are based on the geometry of the regions of interest and the transient dose distributions occurring during the optimization process. These structures are automatically assigned dose-volume objectives and assist in meeting clinical goals [37]. This solution has also shown promising results in the head and neck region [37,38].

(18)

15

1

such geometrical information largely determines the degree of achievable sparing for an individual patient, thus one may use these relationships to predict the dose for a new patient geometry. Several groups have investigated the relationship between OAR-PTV geometries and achievable dosimetry, demonstrating its use in quality assurance and dose-predictions for a new patient [39–45]. The work of Appenzoller et al. [46] and Yuan et al. [26], in particular, largely contributed to

the commercial knowledge-based planning solution RapidPlanTM_{(Varian Medical}

Systems, Palo Alto, CA). RapidPlan utilizes a model-library comprising treatment plans of previously treated patients. This model can be used to predict a range of achievable dose-volume histograms for the OARs of a new patient. Automation comes in the form of automatic placement of dose-volume objectives on the lower boundary of these prediction ranges, guiding the optimization process.

The variability and time-intensive nature of manual interactive treatment planning can impair objective and efficient selection between treatment modalities for a prospective patient. This is especially relevant for proton therapy, where high costs and limited availability have attracted considerable attention, and have led to a feeling that the decision to use proton therapy (over photons) needs to be justified [47]. The comparison of proton-photon plans will be influenced by the quality of the two plans. Since there is an inherent learning curve attached to new technologies, protons included [48], this may further compound the variation in manual treatment planning and potentially compromise the plan comparison. High-quality proton and photon treatment planning is crucial to patient selection for proton therapy in the Netherlands, which has adopted a model-based approach using NTCP modelling to predict if clinically relevant toxicity reductions are expected to be provided by protons [49].

Outline of this thesis

In Chapter 2 an investigation into the performance of automated treatment

planning using RapidPlan was carried out in the context of HNC. Three separate model-libraries were used to examine the effect of model composition and size on resulting plan quality. Furthermore, two evaluation groups, one comprised of recently treated patients and one comprised of patients treated at the commencement of RapidArc delivery in the VUmc, were utilized to benchmark RapidPlan and investigate changes in treatment planning between these two time periods.

The RapidPlan model library is the source of the “knowledge” referred to in “knowledge-based planning”. It typically contains treatment plans created by planners with varying degrees of skill. Such variation may result in the inclusion

(19)

16

including dosimetric outliers in a HNC model library on resulting plan quality was investigated by the deliberate inclusion of varying numbers of sub-optimal plans in a model library.

Stereotactic body radiotherapy (SBRT) is a guideline recommended treatment, delivering high doses of radiation in only a few fractions for patients with medically inoperable, peripherally located, early-stage non-small cell lung cancer and for selected patients with lung metastases. Such peripheral lesions present varying

tumor size, location and overlap with OARs such as the thoracic wall. In Chapter 4

we investigated whether RapidPlan could be used to create risk-adapted SBRT plans for patients with peripheral lung lesions. Such models could contribute to the clinical implementation of lung SBRT and help simplify the planning process.

Human contouring practices strive to be anatomically faithful and consistent with expert recommended protocols. However, such practices in HNC are

time-consuming and prone to variation. Chapter 5 compares the use of a commercial

automated contouring solution, a much simplified contouring solution, and conventional contouring practices. Along with the geometrical differences of such contours, the difference in dosimetry when using these automated/simplified approaches is examined.

To ensure appropriate use of protons, patient selection should be objective and

efficient. In Chapter 6 the principle of using RapidPlan (intended for photons) to

select patients for either proton or photon therapy is assessed. Two models, one consisting of proton plans and the other of photon plans, were used to predict which therapy each of 10 HNC patients should receive. Subsequent to this investigation, it was concluded that there was sufficient merit in creating a dedicated version of RapidPlan for protons. As a result, the VUmc assisted Varian Medical Systems in developing RapidPlanPT, a dedicated commercial knowledge-based planning solution for proton therapy. Since alterations were made to RapidPlan to create RapidPlanPT

including the accurate modelling of a proton dose distribution, Chapter 7 details and

evaluates RapidPlanPT for HNC using a patient evaluation group originating from the VUmc. However, whilst the VUmc is experienced in RapidArc treatments and thus clinical treatment plans can be deemed suitable for the benchmarking of photon KBPs, it is not a proton center. Therefore, to suitably benchmark the quality of the

VUmc RapidPlanPT model, Chapter 8 involved benchmarking the plans against

three international proton centers. A RapidPlanPT model, comprising plans from the VUmc, was used to create IMPT KBPs for 7 patients from each of the external centers. The three external centers also provided manual plans for the same 7 patients which were compared with their respective KBPs.

Chapter 9 concludes the body of work in this thesis with a discussion that

(20)

17

1

References

[1] Abstract of the Bill of Mortality for the Town of Boston. New Engl J Med Surg Collat Branches Sci

1813;2:316. doi:10.1056/NEJM181307010020327.

[2] Stewart BW, Wild CP. IARC World Cancer Report 2014 2014:19; 403-412. doi:10.1017/ S0020860400079146.

[3] Gatta G, Botta L, Sanchez MJ, et al. Prognoses and improvement for head and neck cancers

diagnosed in Europe in early 2000s: The EUROCARE-5 population-based study. Eur J Cancer 2015;51:2130–43. doi:10.1016/j.ejca.2015.07.043.

[4] Lubin JH, Purdue M, Kelsey K, et al. Total Exposure and Exposure Rate Effects for Alcohol and

Smoking and Risk of Head and Neck Cancer: A Pooled Analysis of Case-Control Studies. Am J Epidemiol 2009;170:937–47. doi:10.1093/aje/kwp222.

[5] Mehanna H, Beech T, Nicholson T, et al. Prevalence of human papillomavirus in oropharyngeal

and nonoropharyngeal head and neck cancer-systematic review and meta-analysis of trends by time and region. Head Neck 2013;35:747–55. doi:10.1002/hed.22015.

[6] Molina JR, Yang P, Cassivi SD, Schild SE, Adjei AA. Non-small cell lung cancer: epidemiology, risk

factors, treatment, and survivorship. Mayo Clin Proc 2008;83:584–94. doi:10.4065/83.5.584.

[7] Nederland Integraal Kankercentrum. Cijfers over kanker n.d. https://www.cijfersoverkanker.nl

(accessed September 2, 2018).

[8] American Cancer Society. Non-Small Cell Lung Cancer Survival Rates, by Stage 2017. https://

www.cancer.org/cancer/non-small-cell-lung-cancer/detection-diagnosis-staging/survival-rates.html#written_by (accessed August 28, 2018).

[9] Borras JM, Lievens Y, Dunscombe P, et al. The optimal utilization proportion of external beam

radiotherapy in European countries: An ESTRO-HERO analysis. Radiother Oncol 2015;116:38–44. doi:10.1016/j.radonc.2015.04.018.

[10] Dijkema T, Raaijmakers CPJ, Ten Haken RK, et al. Parotid gland function after radiotherapy: the combined michigan and utrecht experience. Int J Radiat Oncol Biol Phys 2010;78:449–53. doi:10.1016/j.ijrobp.2009.07.1708.

[11] Ong CL, Palma D, Verbakel WFAR, Slotman BJ, Senan S. Treatment of large stage I-II lung tumors using stereotactic body radiotherapy (SBRT): Planning considerations and early toxicity. Radiother Oncol 2010;97:431–6. doi:10.1016/j.radonc.2010.10.003.

[12] Mittal BB, Purdy JA, Ang KK. Advances in Radiation Therapy. Springer; 1998.

[13] Bortfeld T. IMRT: A review and preview. Phys Med Biol 2006;51:363–79. doi:10.1088/0031-9155/51/13/R21.

[14] Otto K. Volumetric modulated arc therapy: IMRT in a single gantry arc. Med Phys 2008;35:310–7. doi:10.1118/1.2818738.

[15] Verbakel WF a R, Cuijpers JP, Hoffmans D, et al. Volumetric intensity-modulated arc therapy vs. conventional IMRT in head-and-neck cancer: a comparative planning and dosimetric study. Int J Radiat Oncol Biol Phys 2009;74:252–9. doi:10.1016/j.ijrobp.2008.12.033.

[16] Particle Therapy Co-Operative Group. Particle therapy facilities in operation n.d. https://www. ptcog.ch/index.php/facilities-in-operation (accessed July 4, 2018).

[17] Allen AM, Pawlicki T, Dong L, et al. An evidence based review of proton beam therapy: The report of ASTRO’s emerging technology committee. Radiother Oncol 2012;103:8–11. doi:10.1016/j. radonc.2012.02.001.

[18] Brada M, Pijls-Johannesma M, De Ruysscher D. Proton Therapy in Clinical Practice: Current Clinical Evidence. J Clin Oncol 2007;25:965–70. doi:10.1200/JCO.2006.10.0131.

(21)

18

[20] Mohan R, Das IJ, Ling CC. Empowering Intensity Modulated Proton Therapy Through Physics and Technology : An Overview. Radiat Oncol Biol 2017;99:304–16. doi:10.1016/j.ijrobp.2017.05.005. [21] Webb S. The physical basis of IMRT and inverse planning. Br J Radiol 2003;76:678–89. doi:10.1259/

bjr/65676879.

[22] Loo SW, Martin WMC, Smith P, Cherian S, Roques TW. Interobserver variation in parotid gland delineation: A study of its impact on intensity-modulated radiotherapy solutions with a systematic review of the literature. Br J Radiol 2012;85:1070–7. doi:10.1259/bjr/32038456. [23] Walker G V., Awan M, Tao R, et al. Prospective randomized double-blind study of atlas-based

organ-at-risk autosegmentation-assisted radiation planning in head and neck cancer. Radiother Oncol 2014;112:321–5. doi:10.1016/j.radonc.2014.08.028.

[24] Teguh DN, Levendag PC, Voet PWJ, et al. Clinical validation of atlas-based auto-segmentation of multiple target volumes and normal tissue (swallowing/mastication) structures in the head and neck. Int J Radiat Oncol Biol Phys 2011;81:950–7. doi:10.1016/j.ijrobp.2010.07.009.

[25] Daisne J-F, Blumhofer A. Atlas-based automatic segmentation of head and neck organs at risk and nodal target volumes: a clinical validation. Radiat Oncol 2013;8:154. doi:10.1186/1748-717X-8-154.

[26] Yuan L, Ge Y, Lee WR, et al. Quantitative analysis of the factors which affect the interpatient organ-at-risk dose sparing variation in IMRT plans. Med Phys 2012;39:6868–78. doi:10.1118/1.4757927. [27] 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;2:296–305. doi:10.1016/j.prro.2011.11.012.

[28] Peters LJ, O’Sullivan B, Giralt J, et al. Critical impact of radiotherapy protocol compliance and quality in the treatment of advanced head and neck cancer: Results from TROG 02.02. J Clin Oncol 2010;28:2996–3001. doi:10.1200/JCO.2009.27.4498.

[29] Craft DL, Hong TS, Shih HA, Bortfeld TR. Improved planning time and plan quality through multicriteria optimization for intensity-modulated radiotherapy. Int J Radiat Oncol Biol Phys 2012;82:83–90. doi:10.1016/j.ijrobp.2010.12.007.

[30] Craft DL, Halabi TF, Shih H a., Bortfeld TR. Approximating convex pareto surfaces in multiobjective radiotherapy planning. Med Phys 2006;33:3399–407. doi:10.1118/1.2335486.

[31] Breedveld S, Storchi PRM, Voet PWJ, Heijmen BJM. ICycle: Integrated, multicriterial beam angle, and profile optimization for generation of coplanar and noncoplanar IMRT plans. Med Phys 2012;39:951–63. doi:10.1118/1.3676689.

[32] Voet PWJ, Dirkx MLP, Breedveld S, et al. Toward Fully Automated Multicriterial Plan Generation: A Prospective Clinical Study. Int J Radiat Oncol 2013;85:866–72. doi:10.1016/j.ijrobp.2012.04.015. [33] Voet PWJ, Dirkx MLP, Breedveld S, et al. Fully Automated Volumetric Modulated Arc Therapy

Plan Generation for Prostate Cancer Patients. Int J Radiat Oncol 2014;88:1175–9. doi:10.1016/j. ijrobp.2013.12.046.

[34] McGarry CK, Bokrantz R, O’Sullivan JM, Hounsell AR. Advantages and limitations of navigation-based multicriteria optimization (MCO) for localized prostate cancer IMRT planning. Med Dosim 2014;39:205–11. doi:10.1016/j.meddos.2014.02.002.

[35] Tol JP, Dahele M, Peltola J, et al. Automatic interactive optimization for volumetric modulated arc therapy planning. Radiat Oncol 2015;10:75. doi:10.1186/s13014-015-0388-6.

[36] Tol JP, Dahele M, Delaney AR, et al. Detailed evaluation of an automated approach to interactive optimization for volumetric modulated arc therapy plans. Med Phys 2016;43. doi:10.1118/1.4944063.

[37] Gintz D, Latifi K, Caudell J, et al. Initial evaluation of automated treatment planning software. J Appl Clin Med Phys 2016;17:331–46. doi:10.1120/jacmp.v17i3.6167.

(22)

19

1

[39] Wu B, Ricchetti F, Sanguineti G, et al. Patient geometry-driven information retrieval for IMRT

treatment plan quality control. Med Phys 2009;36:5497–505. doi:10.1118/1.3253464.

[40] Zhu X, Ge Y, Li T, et al. A planning quality evaluation tool for prostate adaptive IMRT based on machine learning. Med Phys 2011;38:719. doi:10.1118/1.3539749.

[41] Chanyavanich V, Das SK, Lee WR, Lo JY. Knowledge-based IMRT treatment planning for prostate cancer. Med Phys 2011;38:2515. doi:10.1118/1.3574874.

[42] Yang Y, Ford EC, Wu B, et al. An overlap-volume-histogram based method for rectal dose prediction and automated treatment planning in the external beam prostate radiotherapy following hydrogel injection. Med Phys 2013;40:11709. doi:10.1118/1.4769424.

[43] Wu B, McNutt T, Zahurak M, et al. Fully automated simultaneous integrated boosted-intensity modulated radiation therapy treatment planning is feasible for head-and-neck cancer: a prospective clinical study. Int J Radiat Oncol Biol Phys 2012;84:e647-53. doi:10.1016/j. ijrobp.2012.06.047.

[44] Wu B, Pang D, Simari P, et al. Using overlap volume histogram and IMRT plan data to guide and automate VMAT planning: A head-and-neck case study. Med Phys 2013;40:21714. doi:10.1118/1.4788671.

[45] Lian J, Yuan L, Ge Y, et al. Modeling the dosimetry of organ-at-risk in head and neck IMRT planning: an intertechnique and interinstitutional study. Med Phys 2013;40:121704. doi:10.1118/1.4828788.

[46] Appenzoller LM, Michalski JM, Thorstad WL, Mutic S, Moore KL. Predicting dose-volume histograms for organs-at-risk in IMRT planning. Med Phys 2012;39:7446–61. doi:10.1118/1.4761864. [47] Furlow B. Dosimetric promise versus cost: critics question proton therapy. Lancet Oncol

2013;14:805–6. doi:10.1016/S1470-2045(13)70314-0.

[48] Liao Z, Lee JJ, Komaki R, et al. Bayesian Adaptive Randomization Trial of Passive Scattering Proton Therapy and Intensity-Modulated Photon Radiotherapy for Locally Advanced Non–Small-Cell Lung Cancer. J Clin Oncol 2018:JCO.2017.74.072. doi:10.1200/JCO.2017.74.0720.

(23)

(24)

21

Evaluation of a

Knowledge-Based Planning Solution for

Head and Neck Cancer

Jim P. Tol, Alexander R. Delaney, Max Dahele, Ben J. Slotman, and Wilko F.A.R. Verbakel

International Journal of Radiation Oncology Biology Physics. 2015;91(3):612-20.

Chapter 2

(25)

22

Abstract

Purpose

Automated and knowledge-based planning techniques aim to reduce variations in plan quality. RapidPlan uses a library consisting of different patient plans to make a model that can predict achievable dose-volume histograms (DVHs) for new patients and uses those models for setting optimization objectives. We benchmarked RapidPlan versus clinical plans for 2 patient groups, using 3 different libraries.

Methods and materials

Volumetric modulated arc therapy plans of 60 recent head and neck cancer patients that included sparing of the salivary glands, swallowing muscles, and oral

cavity were evenly divided between 2 models, Model_30A and Model_30B, and were

combined in a third model, Model₆₀. Knowledge-based plans were created for 2

evaluation groups: evaluation group 1 (EG1), consisting of 15 recent patients, and evaluation group 2 (EG2), consisting of 15 older patients in whom only the salivary glands were spared. RapidPlan results were compared with clinical plans (CP) for boost and/or elective planning target volume homogeneity index, using

HI_B/HI_E = 100 × (D2% − D98%)/D50%, and mean dose to composite salivary

glands, swallowing muscles, and oral cavity (D_sal, D_swal, and D_oc, respectively).

Results

For EG1, RapidPlan improved HI_B and HI_E values compared with CP by 1.0%

to 1.3% and 1.0% to 0.6%, respectively. Comparable D_sal and D_swal values were

seen in Model_30A, Model_30B, and Model₆₀, decreasing by an average of 0.1, 1.0,

and 0.8 Gy and 4.8, 3.7, and 4.4 Gy, respectively. However, differences were

noted between individual organs at risk (OARs), with Model_30B increasing D_oc by

0.1, 3.2, and 2.8 Gy compared with CP, Model_30A, and Model₆₀. Plan quality was

less consistent when the patient was flagged as an outlier. For EG2, RapidPlan

decreased D_sal by 4.1 to 4.9 Gy on average, whereas HI_B and HI_E decreased by

1.1% to 1.5% and 2.3% to 1.9%, respectively.

Conclusions

(26)

23

2

Introduction

Variation in knowledge and experience can lead to large differences in the quality of radiation therapy treatment plans [1, 2] and may compromise the gains that can be realized with advanced technologies such as intensity modulated radiation therapy (IMRT) and volumetric modulated arc therapy (VMAT). The same holds true for labor and computing resources, which can affect the implementation of new treatment planning techniques and treatment planning capacity. Various solutions are being investigated to improve planning consistency [3-7], including increased automation of planning by using knowledge-based approaches [8-13]. These approaches typically use libraries of existing patient plans to create models that predict the amount of organ-at-risk (OAR) sparing that can be achieved for a new patient, based, for example, on planning target volume (PTV)-OAR distance and overlap [14]. Being able to rationally predict OAR dose-volume histograms (DVHs) could remove the need for performing interactive optimization or multiple iterative optimizations and could increase consistency in treatment planning [14-18]. Resulting plans produced by such a knowledge-based system should reflect the quality of the plans that populate the model and the ability of the software to predict the achievable DVHs.

RapidPlan (Varian Medical Systems, Palo Alto, CA) is a commercially available knowledge-based planning solution derived from previously published work [14, 16]. Knowledge in this case is represented by models created from libraries of previous plans. The purpose of this report was (1) to benchmark RapidPlan performance compared to recent clinical VMAT (RapidArc, Varian) plans by using model libraries made up of different plans and with different total numbers of plans; and (2) to investigate whether model libraries based on plans that spare many OARs can be usefully applied to patients treated shortly after our clinical introduction of RapidArc, in whom fewer OARs were spared. This allowed us to test the versatility of the model libraries and to illustrate what a center new to RapidArc planning and starting with the inclusion of only a few OARs may anticipate when applying a model consisting of more advanced plans.

Methods and Materials

Clinical plans

(27)

24

of the elective PTV (PTV_E; V95 ≥ 98%) and 95% of the prescribed dose of 70 Gy

to 99% of the boost PTV (PTV_B), in 35 fractions, while limiting the volume of

each PTV receiving >107% (V107) of the prescribed dose. A 5-mm transition zone

(PTV_T) was created between PTV_E and PTV_B to facilitate dose falloff between

them. OAR planning goals included maximizing point doses to the spinal cord, brainstem, and their planning at risk volumes (3-mm expansion), and lowering the mean dose to parotid and submandibular glands (SMGs), individual swallowing muscles, and the oral cavity as much as possible. RapidArc optimization for HNC was performed interactively by continuously adapting the location of the OAR optimization objectives according to our institutional optimization protocol [19-21]. Reduced clinical plan quality could occasionally be obtained if the institutional optimization protocol was not followed thoroughly.

Model libraries

RapidPlan uses model libraries that contain dose distributions and OAR and PTV geometries of previously treated patients to generate a prediction range of achievable DVHs for individual OARs of new patients [22, 23]. Optimization is automated by placing numerous dose-volume objectives along the lower range of the predicted DVHs. Although RapidPlan can calculate optimal priorities for optimization objectives, this feature is still being refined. A set of priorities reflecting our institutional practice was therefore entered manually. Clinical RapidArc plans of 60 HNC patients treated between 2012 and 2014 (using Eclipse treatment planning system version 10.0.28; Varian) were arbitrarily selected and

evenly divided among 2 models, Model_30A and Model_30B, to investigate whether

differences in the composition of plan libraries influenced RapidPlan results.

An additional model consisting of all 60 plans (Model₆₀) was used to evaluate

the influence of model size. These models were used to create RapidArc plans for 2 evaluation groups, using Eclipse version 13.5 software, with the photon optimizer (PO) algorithm version 13.5.10. These plans all used the same normal tissue objective settings. Dose calculation, with a subsequent “continue previous optimization” (CPO) calculation to improve PTV dose homogeneity [19], was performed using the anisotropic analytical algorithm (AAA) version 13.0.16 with a 2.5-mm calculation grid. The knowledge-based plans were normalized to deliver

the same mean dose to PTV_B as the respective clinical plan (CP).

(28)

25

2

in the CP were added to the model library. This meant that in some cases, fewer than 20 individual OARs were available from the 30 plans. As a result, there were inadequate data with which to generate model estimations of individual superior, medial, and inferior pharyngeal constrictor muscles (PCM) and upper and lower larynx structures. To circumvent this, composite PCM and larynx structures were created from relevant, individually spared OARs to estimate DVHs for individual PCM and larynx structures. Both of these composite structures were present in at least 20 plans in all models. Similarly, contralateral SMGs also were used to model ipsilateral SMGs.

Evaluation groups

Evaluation group 1 (EG1) consisted of clinical treatment plans from 15 patients treated between 2013 and 2014, similar to those included in the models. EG1 was used to benchmark RapidPlan results compared to recent CPs. Optimization and dose calculation were performed using the progressive resolution optimizer (PRO) and AAA version 10.0.28. OARs typically included the oral cavity, salivary glands, and swallowing muscles. Depending on the degree of OAR-PTV overlap and whether the treating clinician chose to spare them, the salivary OARs could consist of some or all of the contra- and ipsilateral parotid and SMGs. Similarly, swallowing OARs could consist of some or all of the upper esophageal sphincter, upper and lower larynx, superior, medial and inferior PCMs, cricopharyngeal muscle, and esophagus.

Evaluation group 2 (EG2) consisted of plans from 15 patients treated in 2008 and 2009, with plans developed using PRO or AAA versions 8.2.23 to 8.6.15. These plans were made at the beginning of our department’s RapidArc program by a relatively inexperienced team. Attempts were made to spare the parotid and, less frequently, the submandibular glands, whereas the oral cavity and swallowing muscles were not spared.

Plans included in EG1 and EG2 were not included in the RapidPlan model libraries. To simplify OAR dose reporting, composite salivary and swallowing

structures (comp_sal and comp_swal, respectively), consisting of individually

constrained OARs, were created. The size of the PTVs, comp_sal, comp_swal, and oral

cavities included in the 3 models and 2 evaluation groups are shown in Table 1, and Table 2 shows the number of OARs included.

(29)

26

of 5 patients from the 2008-2009 treatment period were recreated using PRO and AAA version 10.0.28, PO version 13.5.10, and AAA version 13.0, without altering the optimization objectives.

Table 1.

Volume (cm3₎ _PTV

Ba PTVEa PTVTa Compsalb Compswalc Oral cavity

Model_30A Mean 208.1 ± 119.6 346.9 ± 74.8 74.0 ± 34.0 66.2 ± 22.7 22.6 ± 15.1 70.1 ± 33.8 Range 39.1 to 607.0 223.5 to 514.2 17.6 to 155.7 20.7 to 118.2 4.8 to 67.5 14.7 to 186.6 Model_30B Mean 150.4 ± 102.1 390.3 ± 102.4 60.8 ± 51.3 65.8 ± 19.4 25.4 ± 14.1 108.2 ± 68.0 Range 34.1 to 536.1 240.6 to 618.6 10.1 to 258.4 22.6 to 101.7 3.4 to 57.7 29.4 to 283.5 Model₆₀ Mean 179.3 ± 114.0 368.6 ± 91.6 67.7 ± 43.2 66.0 ± 21.0 24.0 ± 14.6 88.4 ± 55.9 Range 34.1 to 607.0 223.5 to 618.6 10.1 to 258.4 20.7 to 118.2 3.4 to 67.5 14.7 to 283.5 EG1d Mean 236.7 ± 150.2 376.8 ± 108.3 89.0 ± 45.2 67.3 ± 27.5 27.0 ± 11.1 87.1 ± 55.3 Range 69.9 to 666.7 207.6 to 658.6 27.0 to 199.1 23.6 to 105.7 5.1 to 46.3 35.1 to 241.9 EG2e Mean 200.6 ± 108.5 367.1 ± 121.7 51.6 ± 37.8 57.0 ± 18.8 - -Range 75.5 to 475.6 226.2 to 657.4 15.9 to 156.1 29.8 to 97.1 -

-a _{Boost, elective and transition planning target volumes}

b _{Composite salivary glands}

c _{Composite swallowing muscles}

d _{The first evaluation group, consisting of 15 patients treated at our institute between 2012 and 2014}

e_{The second evaluation group, consisting of 15 patients treated at our institute in 2008 or 2009}

Size of the planning target volumes (PTVs), composite salivary / swallowing structures (comp_sal /

compswal) and oral cavities of the patients included in the three RapidPlan models (Model30A, Model30B

and Model₆₀) and two evaluation groups (EG1 and EG2).

Outlier detection

(30)

27

2

Table 2.

Model_30A Model_30B Model₆₀ EG1d _EG2e

Oral Cavity 27 24 51 15 -CL parotida ₃₀ ₃₀ ₆₀ ₁₅ ₁₅ IL parotida ₂₇ ₂₉ ₅₆ ₁₅ ₁₅ CL SMGb ₂₅ ₂₅ ₅₀ ₁₀ ₆ IL SMGb ₅ ₅ ₁₀ ₂ ₄ Cricopharyngeal Muscle 24 23 47 11 -Lower Larynx 19 20 39 12 -Upper Larynx 14 14 28 9 -Inferior PCMc ₁₉ ₁₉ ₃₈ ₉ -Medial PCMc ₁₄ ₁₃ ₂₇ ₄ -Superior PCMc ₁₄ ₁₈ ₃₂ ₆

-Upper esophageal sphincter 30 26 56 14

-a_{Contralateral and ipsilateral parotid glands}

b_{Contralateral and ipsilateral submandibular glands}

c_{Pharyngeal constrictor muscles}

d_{The first evaluation group, consisting of 15 patients treated at our institute between 2012 and 2014}

e_{The second evaluation group, consisting of 15 patients treated at our institute in 2008 or 2009}

Number of organs at risk included in RapidPlan Model_30A, Model_30B, and Model₆₀ and evaluation

groups EG1 and EG2

Principal component analysis is used to determine the geometric feature that correlates most strongly with the spread of dosimetry in the model plans. This can allow the user to reject certain plans (eg those in which insufficient OAR sparing was achieved) to try and improve the model. This warning system may be further refined in future versions, and it should be used in conjunction with visual and dosimetric plan evaluation when deciding whether to reject a plan from the

model. Using this strategy, 6 of 60 patients in Model₆₀ were ultimately identified

as containing 1 outlier OAR. Nevertheless, these patients were included in the models to investigate RapidPlan’s performance, providing a set of arbitrarily selected clinical plans without removing or replanning of outlier plans.

Before applying a model to calculate OAR DVH estimations for a patient, RapidPlan evaluates whether the OAR is an outlier according to a number of model parameters or whether the OAR’s geometry is located in a poorly populated region of the model. The evaluation parameters are OAR and PTV (combined as

PTV_B + PTV_E + PTV_T) volumes, OAR-PTV distance and overlap, and the first

(31)

28

alert the user to the possibility of unreliable DVH estimations that might lead to suboptimal plans.

Study endpoints

A comparison between the knowledge-based plans and their respective CPs was

made on the basis of (1) the homogeneity index (HI) calculated for PTV_B/PTV_E (HI_B/

HI_E) using [HI = 100% x (D2%-D98%)/D50%]; and (2) mean doses to individual

OARs, comp_sal, comp_swal, and oral cavity (D_sal, D_swal, and D_OC, respectively). Paired,

2-sided Student t-tests were performed to identify significant (P≤.05) differences between RapidPlan and clinical plans.

Results

In EG1, an average of 1.8, 1.1, and 0.5 OARs per patient were flagged as outliers

by Model_30A, Model_30B, and Model₆₀, respectively, from an average of 8.3 OARs

per patient. In EG2, there were 0.4, 0.5, and 0.3 outlier OARs, respectively, per

patient of 2.7 OARs per patient. Model₆₀resulted in the least number of outliers,

indicating that this model can account for a larger range of patient geometries. Although we analyzed the outlier warnings given when applying the different models, investigating the influence of outliers on RapidPlan performance is beyond the scope of this study.

All knowledge-based plans were reviewed and deemed satisfactory by a senior clinical physicist experienced in HNC planning. Table 3 summarizes RapidPlan results for EG1, averaged over all patients, and Figure 1 shows results for

individual patients. Compared with the CPs, PTV_B coverage (V95), V107, and

HI_B/HI_E values improved with the use of RapidPlan, although most differences

were not significant. On average, all 3 models provided comparable D_sal and D_swal

values. Compared to the CP, D_swal decreased by 3.7 to 4.8 Gy using RapidPlan.

This was caused mostly by an average of 4.3 to 6.0 Gy more sparing of the upper and lower larynx and inferior PCM. Differences between the models were noted

for some individual OARs. For example, Model_30B resulted in an average D_OC of

26.5 Gy; that is, 0.1, 3.2, and 2.8 Gy higher than in CP, Model_30A, and Model₆₀,

respectively. The following examples of EG1 patient outliers are highlighted.

In patient 11 (Fig. 1), who had a large PTV_B volume (666.7 cm3_{), the combined}

target volumes (PTV_B + PTV_E + PTV_T = 1137.9 cm3_{) and OAR-PTV overlap}

volume were flagged as outliers for all 8 OARs, using Model_30A, whereas only the

(32)

29

2

and Model₆₀. D_sal, D_swal, and D_OC results with Model_30A, Model_30B, Model₆₀ were

41.0, 44.3, 44.5, 21.6, 29.2, and 23.6 Gy and 30.6, 50.1, and 28.3 Gy, respectively. In patients 6 and 3 (Fig. 1), all included OARs were flagged as target volume

outliers by using Model_30A, whereas Model_30B and Model₆₀ flagged 25% and 25%

and 0% and 0% of OARs as outliers in these patients, respectively. In patient

6, highly variable plan quality was obtained, with maximum differences in D_sal,

D_swal, and D_OC of 1.8, 8.0, and 11.3 Gy, respectively, between models. The worst

performing model, Model_30B, increased D_OC by 8.1 Gy compared to that in CP. In

contrast, D_sal, D_swal, and D_OC varied in patient 3 by 3.1, 1.3, and 4.4 Gy, respectively,

between models, and the gains compared to those with CP were similar to those demonstrated by the pooled data.

A number of findings illustrate different trade-offs in the knowledge-based

planning results. In patient 5, the models resulted in decreased D_OC (3.8-7.0 Gy)

and D_sal(3.6-4.6 Gy) at the cost of higher HI_B (9.6%-10.4% vs 8.4% clinically)

and HI_E (12.8%-13.4% vs 11.2%). In patient 10, Model_30A improved D_OC by 7.8

Gy compared to the clinical plan, at the cost of 2.3-Gy-higher D_sal. In contrast,

Model_30B and Model₆₀ decreased D_OC by 0.7 and 4.5 Gy, respectively, whereas D_sal

was reduced by 0.6 and 1.3 Gy, respectively. Figure 2 shows dose distributions and DVH results for patient 1, in whom the plan made by RapidPlan resulted

in substantially improved comp_swalsparing compared to that by the CP but also

decreased dose conformity outside the PTVs. Visual inspection of the CP revealed deviation from the planning protocol with too few optimization objectives, suboptimally located objectives for the individual swallowing muscles. The same was true for patient 14.

Results obtained for EG2, averaged over all 15 patients, are summarized in

Table 4. Pooled data showed significant improvements in PTV_B coverage and

sparing of both parotid glands using RapidPlan. Compared to the CPs, RapidPlan

decreased HI_B and HI_E by 1.1% to 1.5% and 1.9% to 2.3%, respectively, whereas

D_sal decreased (significantly) by 4.1 Gy to 4.9 Gy. In 5 patients, all 3 models

achieved >6 Gy reduction in D_sal compared to the CP.

The effect of different optimization and dose calculation algorithms on OAR and PTV dosimetry (averaged for 5 patients) is summarized in Table 5. Newer optimization algorithms primarily improved PTV dose homogeneity, whereas OAR sparing was less influenced.

After importing all plans and assigning the OAR and PTV structures, computing

time to create Model_30A, Model_30B, and Model₆₀ required 19.1, 19.5, and 39.2

(33)

30

Table 3.

Plan Clinical Model30A

Model_30A (Range) Model30B Model_30B (Range) Model60 Model₆₀ (Range) Boost PTV (%) V95 99.0 ± 0.0 99.3 ± 0.6 -1.0 to 1.1 99.5 ± 0.4* -1.0 to 0.7 99.5 ± 0.5* -1.0 to 0.8 V107 2.6 ± 3.9 2.0 ± 2.5 -6.7 to 7.9 1.4 ± 2.2 -1.4 to 7.8 1.5 ± 2.1 -1.8 to 7.7 HIa _{10.3 ± 1.2} _{9.3 ± 1.1} _{-2.0 to 4.5} _{9.0 ± 1.3} _{-1.2 to 5.2} _{9.2 ± 1.3} _{-1.6 to 4.7} Elective PTV (%) V95 97.9 ± 0.8 97.8 ± 1.5 -2.3 to 2.9 98.2 ± 1.3 -2.7 to 2.7 98.1 ± 1.0 -2.2 to 1.7 V107 13.6 ± 9.0 11.2 ± 4.2 -10.3 to 22.6 9.7 ± 3.8 -4.6 to 24.6 9.9 ± 4.0 -4.0 to 23.4 HIa _{15.4 ± 2.3} _{14.7 ± 1.7} _{-2.6 to 4.2} _{14.4 ± 1.7} _{-2.2 to 4.6} _{14.8 ± 1.6} _{-2.9 to 4.2}

Max dose (Gy)

Spinal Cord 40.5 ± 5.1 40.4 ± 0.7 -6.9 to 1.3 40.3 ± 0.8 -6.8 to 0.7 40.4 ± 0.9 -7.1 to 1.2

Brainstem 40.1 ± 8.0 38.2 ± 3.8 -6.5 to 14.0 38.8 ± 2.2 -2.8 to 13.9 37.9 ± 3.5 -7.0 to 13.2

Mean dose (Gy)

Oral Cavity 26.4 ± 11.2 23.3 ± 8.9* -2.3 to 8.3 26.5 ± 10.8 -11.2 to 9.3 23.7 ± 9.8* -0.7 to 7.4 CL Parotidb _{20.3 ± 4.3} _{20.0 ± 4.1} _{-3.1 to 2.5} _{18.4 ± 4.2*} _{-3.4 to 4.4} _{19.1 ± 3.6*} _{-1.2 to 3.9} IL Parotidb _{27.0 ± 5.7} _{27.2 ± 6.7} _{-3.2 to 7.5} _{27.1 ± 6.9} _{-4.2 to 5.9} _{26.9 ± 7.1} _{-6.1 to 5.4} CL SMGc _{33.2 ± 5.1} _{33.1 ± 4.5} _{-4.0 to 4.3} _{31.2 ± 6.8} _{-1.1 to 9.3} _{31.3 ± 6.3} _{-1.5 to 6.9} IL SMGc _{52.3 ± 8.5} _{46.8 ± 9.3} _{4.9 to 6.1} _{47.7 ± 10.9} _{2.8 to 6.3} _{47.2 ± 10.8} _{3.5 to 6.8} Cricophd _{27.6 ± 10.5} _{26.1 ± 11.0} _{-7.0 to 15.8} _{25.8 ± 9.1} _{-8.9 to 15.9} _{25.5 ± 9.8} _{-4.5 to 18.5} Lower Larynx 27.5 ± 9.0 23.1 ± 9.3 -3.7 to 21.7 22.6 ± 8.7* -1.3 to 20.0 22.5 ± 9.6* -3.3 to 20.0 Upper Larynx 40.0 ± 12.5 34.5 ± 10.8* 0.6 to 12.2 34.7 ± 10.4* -0.4 to 12.3 34.3 ± 10.8* -0.3 to 11.3 Inferior PCMe _{30.9 ± 8.2} _{25.2 ± 4.9*} _{-2.3 to 14.9} _{25.4 ± 4.3*} _{-2.0 to 15.3} _{25.1 ± 4.6*} _{-0.3 to 14.1} Medial PCMe _{52.0 ± 8.1} _{50.6 ± 7.6} _{-1.7 to 3.6} _{51.4 ± 7.0} _{-2.2 to 3.3} _{51.6 ± 6.7} _{-3.9 to 3.5} Superior PCMe _{44.8 ± 9.0} _{40.9 ± 8.7*} _{1.0 to 5.3} _{42.1 ± 9.7*} _{-0.1 to 5.6} _{40.4 ± 9.1*} _{1.4 to 6.2} UESf _{23.3 ± 10.3} _{19.9 ± 7.3} _{-4.0 to 16.2} _{19.6 ± 10.5*} _{-4.5 to 18.3} _{19.8 ± 9.0*} _{-3.6 to 16.7} Comp_salg 24.6 ± 4.3 24.5 ± 4.5 -2.3 to 3.6 23.6 ± 4.3 -2.5 to 4.6 23.8 ± 4.3 -2.5 to 4.6 Compswalg 32.9 ± 7.9 28.1 ± 5.9* -2.7 to 14.2 29.2 ± 6.7* -0.8 to 11.4 28.5 ± 6.3* -1.4 to 11.8 * Statistically significant difference (p≤0.05) with clinical plan result

a boost and elective PTV homogeneity indices b Contralateral and ipsilateral parotid glands c Contralateral and ipsilateral submandibular glands d Cricopharyngeal muscle

e Pharyngeal constrictor muscle f Upper esophageal sphincter

g Composite salivary and swallowing structures

(34)

31

2

Figure 1. Histograms showing mean doses to the contra- and ipsilateral parotid glands, contralateral

(35)

32

Figure 2. Dose distributions and dose-volume histograms of the clinical plan and three

(36)

33

2

Table 4.

Plan Clinical Model_30A Model30A

(Range) Model30B Model_30B (Range) Model60 Model₆₀ (Range) Boost PTV (%) V95 99.3 ± 0.5 99.8 ± 0.2* -1.9 to 0.2 99.8 ± 0.2* -2.0 to 0.3 99.7 ± 0.2* -1.9 to 0.3 V107 1.7 ± 2.9 0.8 ± 1.5 -0.6 to 6.3 1.0 ± 1.6 -0.6 to 6.2 1.2 ± 2.2 -3.0 to 5.4 HIa _{8.7 ± 1.7} _{7.2 ± 1.2*} _{-0.3 to 4.3} _{7.5 ± 1.4*} _{-1.7 to 3.9} _{7.6 ± 1.5} _{-1.3 to 3.5} Elective PTV (%) V95 97.8 ± 3.1 99.3 ± 0.7 -10.5 to 1.1 99.2 ± 0.9* -8.8 to 0.7 99.3 ± 0.5 -10.7 to 0.7 V107 7.9 ± 9.6 4.5 ± 4.0 -3.5 to 22.9 5.9 ± 4.9 -5.7 to 14.7 5.8 ± 6.0 -5.8 to 13.6 HIa _{13.0 ± 4.0} _{10.8 ± 2.3*} _{-1.3 to 8.9} _{11.1 ± 2.8} _{-1.5 to 6.2} _{10.7 ± 2.3*} _{-0.9 to 9.3}

Max dose (Gy)

Spinal Cord 43.4 ± 6.8 43.1 ± 6.2 -5.3 to 3.6 43.7 ± 6.7 -5.5 to 4.0 43.9 ± 6.7 -4.7 to 5.4

Brainstem 31.1 ± 16.5 30.7 ± 16.8 -2.2 to 4.6 30.2 ± 16.7 -1.2 to 5.5 30.9 ± 16.7 -2.5 to 5.1

Mean dose (Gy)

CL Parotidb _{23.9 ± 6.5} _{20.2 ± 6.5*} _{-0.6 to 8.9} _{19.1 ± 5.7*} _{1.3 to 10.7} _{19.5 ± 5.9*} _{-1.1 to 10.9}

IL Parotidb _{31.5 ± 8.8} _{27.4 ± 7.9*} _{-2.6 to 10.4} _{27.1 ± 8.9*} _{-2.0 to 12.2} _{26.5 ± 7.5*} _{-1.4 to 10.8}

CL SMGc _{40.6 ± 11.0} _{36.6 ± 12.4} _{-2.7 to 19.4} _{36.3 ± 12.2} _{-3.0 to 19.6} _{35.4 ± 12.6} _{-3.2 to 19.3}

IL SMGc _{50.1 ± 16.7} _{49.6 ± 19.0} _{-1.0 to 4.3} _{52.0 ± 15.8*} _{-0.9 to -2.8} _{47.0 ± 18.0} _{-1.0 to 8.6}

Compsald _{28.8 ± 6.2} _{24.7 ± 5.7*} _{-1.3 to 8.2} _{24.2 ± 6.0*} _{0.2 to 9.7} _{23.9 ± 5.3*} _{-1.1 to 8.6} * Statistically significant difference (p≤0.05) with clinical plan result

a boost and elective PTV homogeneity indices b Contralateral and ipsilateral parotid glands c Contralateral and ipsilateral submandibular glands d Composite salivary glands

Organ-at-risk (OAR) and planning target volume (PTV) dosimetry obtained when applying Model_30A,

Model_30B and Model₆₀ to the second evaluation group (EG2), consisting of patients treated between

2008-2009. Data is averaged over all 15 patients. The range shows the smallest and largest deviation when computing clinical minus RapidPlan results.

Table 5.

Algorithm PRO & AAA v8.6-8.9 PRO & AAA v10.0.28 PO & AAA v13.0 Homogeneity indices (%)

Boost PTV 9.5 ± 2.2 7.4 ± 1.5 7.5 ± 1.7

Elective PTV 13.1 ± 2.4 10.2 ± 2.3 10.6 ± 2.3

Mean dose (Gy)

CL Parotida _{19.7 ± 1.4} _{19.3 ± 1.8} _{18.9 ± 2.0}

IL parotida _{26.8 ± 5.0} _{26.2 ± 4.6} _{25.7 ± 4.5}

CL SMGb _{37.3 ± 6.8} _{36.9 ± 6.4} _{37.3 ± 7.7}

IL SMGb _{52.1 ± 12.2} _{51.4 ± 12.2} _{51.7 ± 12.4}

Compsalc _{27.5 ± 5.1} _{27.0 ± 5.5} _{26.7 ± 4.8} a: Contralateral and ipsilateral parotid glands

b: Contralateral and ipsilateral submandibular glands c: Composite salivary glands

(37)

34

Discussion

In this study, we evaluated the performance of a commercial knowledge-based planning solution. HNC patients were chosen for the analysis because their plans include multiple PTVs and many individual salivary and swallowing OARs, testing RapidPlan performance in a relatively challenging, although common, clinical scenario. Pooled results showed that generally, if most of the OARs in a patient were not flagged as outliers, RapidPlan provided comparable and often improved plan quality compared to that of CPs. The superiority of certain RapidPlan plans compared with CPs could be due to the challenging nature of optimally and consistently performing interactive planning for plans which contain many OARs, within a limited number of iterations. Pooled data illustrate the fact that models created using 30 plans created plans that were similar to those in a model based on 60 plans. However, these models could produce substantially different results in individual patients for specific OARs. This indicates that plans produced by the models are sensitive to the composition of the plan library and to the characteristics of the patient for which the knowledge-based plan is made. We also observed that more OAR outliers did not necessarily translate into a worse OAR dose.

RapidPlan results for specific OARs in EG1 showed appreciable variations among

the models. For example, Model_30B plans resulted in 3.2 and 2.8 Gy higher D_OC than

Model_30A and Model₆₀, respectively. The following factors may have contributed: (1)

Model_30B included the least number of oral cavities (24 vs 27 and 51 in Model_30A and

Model₆₀); (2) the oral cavity volumes in Model_30B (mean, 108 cm3_{; range, 29-284 cm}3₎

were, on average, larger than in EG1 (mean, 87 cm3_{; range, 35-242 cm}3_{); and (3) in the}

3 patients who showed the largest increase in D_OC with Model_30B compared to the CPs,

the first patient had a large PTV_B (266.5 cm3_{) and combined PTV volume (1044.1 cm}3₎

compared to the model (averages of 150.4 ± 102.1 cm3_{and 594.7 ± 222.4 cm}3_,

respectively). The second patient had a small oral cavity (41.3 cm3_{vs 108.2 cm}3_{in the}

model), and the third patient had both a small oral cavity (39.7 cm3_{) and a large PTV}

B

volume (666.7 cm3_).

Although it is expected that including more patients in the model libraries would lead to fewer outliers and more consistent plan quality, this requires further investigation.

We observed for example that, although Model₆₀ resulted in fewer OAR outliers in

EG1 and EG2, this did not translate into consistently lower OAR doses. For this study, patients were randomly selected for inclusion in the models, and no CPs were reoptimized on the basis of model statistics. The influence of the dosimetric outliers in

Model₆₀ on resulting plan quality was evaluated by removing these outliers from the

model and replanning for the first 5 patients of EG1. The effect was marginal, with HI_B

and HI_E differing by less than 0.0 ± 0.4% and 0.4 ± 0.3%, on average, whereas D_OC,

(38)

35

2

on the OAR-PTV geometry of the plans included in the models, consistent contouring is important in generating reliable model plan libraries. This might be helped by the use of contouring atlases [24]. Note that clinical inspection of plans created using RapidPlan remains necessary.

With respect to EG2, the RapidPlan models achieved substantial gains in sparing of most OARs compared with the CPs. These results indicate the gains in OAR sparing that inexperienced centers starting VMAT treatments might anticipate when using a RapidPlan model created by a more experienced center producing higher quality plans. They also show that a model consisting of plans with many OARs is versatile and can be successfully applied to patients with fewer OARs.

Some of the improvements (eg in PTV dose homogeneity) can be attributed to the newer optimization algorithms available in Eclipse version 13.5. This may be because the CPO functionality that uses the final AAA calculated dose distribution as a starting point for a new optimization was not available in PRO version 8.6 to 8.9, although it has been shown to substantially improve PTV dose homogeneity [19].

Results of several in-house-developed knowledge-based planning solutions have been previously reported. Moore et al [15], evaluated the correlation between the fraction of OARs overlapping with the PTVs and the mean OAR doses and were able to identify patients in whom large gains in parotid gland sparing could be achieved. These gains were validated by replanning for the identified patients. Predicting achievable DVHs by using a library of previously delivered plans has been investigated by various groups. Yuan et al [16] used principal component analysis to identify significant patient anatomical factors contributing to OAR sparing in head and neck IMRT plans and applied this knowledge to realize gains in parotid gland sparing. Appenzoller et al [14] identified differences between predicted and obtained DVHs to identify possible outliers with regard to OAR sparing. Replanning these outliers allowed refinement of the model.