Automated multi-criterial planning with beam angle optimization to establish non-coplanar VMAT class solutions for nasopharyngeal carcinoma

Purpose: Complexity in selecting optimal non-coplanar beam setups and prolonged delivery times may hamper the use of non-coplanar treatments for nasopharyngeal carcinoma (NPC). Automated multi-criterial planning with integrated beam angle optimization was used to define non-coplanar VMAT class solutions (CSs), each consisting of a coplanar arc and additional 1 or 2 fixed, non-coplanar partial arcs. Methods: Automated planning was used to generate a coplanar VMAT plan with 5 complementary computer- optimized non-coplanar IMRT beams (VMAT + 5) for each of the 20 included patients. Subsequently, the frequency distribution of the 100 patient-specific non-coplanar IMRT beam directions was used to select non- coplanar arcs for supplementing coplanar VMAT. A second investigated CS with only one non-coplanar arc consisted of coplanar VMAT plus a partial arc at 90 ◦ couch angle (VMATCS90). Plans generated with the two VMATCSs were compared to coplanar VMAT. Results: VMAT + 5 analysis resulted in VMATCS60: two partial non-coplanar arcs at couch angles 60 ◦ and (cid:0) 60 ◦ to complement coplanar VMAT. Compared to coplanar VMAT, the non-coplanar VMATCS60 and VMATCS90 yielded substantial average dose reductions in OARs associated with xerostomia and dysphagia, i.e., parotids, submandibular glands, oral cavity and swallowing muscles (p < 0.05) for the same PTV coverage and without violating hard constraints. Impact of non-coplanar treatment and superiority of either VMACS60 and VMATCS90 was highly patient dependent. Conclusions: Compared to coplanar VMAT, dose to OARs was substantially reduced with a CS with one or two non-coplanar arcs. Preferences for coplanar or one or two additional arcs are highly patient-specific, balancing plan quality and treatment time.


Introduction
Nasopharyngeal carcinoma (NPC) tumors have planning target volumes (PTV) surrounded by many organs-at-risk (OARs). The standard of care for NPC is coplanar volumetric arc therapy (VMAT) [1], which is known to result in fast treatments with high coverage for irregular targets [2,3], while limiting excessive dose to the surrounding OARs. The most common radiation induced complications are associated excessive dose in the parotid glans, submandibular glands, oral cavity or in the swallowing muscles, which are OARs associated with xerostomia and dysphagia. Studies on several treatment sites have shown that noncoplanar beam arrangements could lead to improved healthy tissue sparing [4][5][6][7][8][9], possibly minimizing side-effects.
Wild et al. [4] compared different coplanar and non-coplanar treatment techniques for nasopharyngeal carcinoma. They found that optimized non-coplanar VMAT treatments reduced mean and maximum doses in the OARs analyzed (eyes, optic nerves and chiasm) compared to coplanar VMAT, while maintaining target coverage. Another method that resulted in improved OAR sparing for NPC was the combination of VMAT and intensity modulated radiotherapy (IMRT) in different fractions for the same patient. Akbas et al. [10] compared IMRT, VMAT and a hybrid IMRT and VMAT technique (each used in different fractions). This hybrid solution resulted in superior conformity for the targets and improved sparing of the brainstem and spinal cord. Some studies have investigated the optimization of dynamic noncoplanar VMAT trajectories, where the gantry and couch rotate simultaneously, as a treatment option for nasopharyngeal carcinoma [4,9]. In both these studies the non-coplanar plans outperformed the coplanar plans, but treatment optimization times were very long. Moreover, these dynamic treatments can currently not be clinically delivered on regular linacs. Instead, optimization times for fixed non-coplanar VMAT arcs are acceptable and clinical application is feasible. To limit treatment times, the use of only few non-coplanar arcs per patient is then preferred.
Selection of optimal patient-specific couch angles for non-coplanar arcs is complex and may be highly workload intensive in case of manual planning. In principle, a fixed arc-class-solution (CS) to be used for all patients could possibly be a solution in clinical planning. However, development of such a CS with manual trial-and-error planning would again be arduous and time-consuming. As far as we are aware of, the use of automated planning for development of VMAT CSs, consisting of a regular coplanar arc supplemented with a few fixed non-coplanar arcs, has never been investigated for NPC.
To improve dose distributions while keeping delivery times acceptable, we previously proposed the VMAT+ treatment approach, complementing coplanar VMAT with ≤ 5 static patient-specific noncoplanar IMRT beams, and tested it for liver SBRT [11] and prostate SBRT [12]. Adding the fully automatically selected non-coplanar IMRT beams improved OAR sparing compared to coplanar VMAT. For prostate, VMAT+5 plans were used to derive a fixed 2-beam non-coplanar CS for the whole patient population to complement coplanar VMAT (VMAT+CS). VMAT+CS resulted in similar plan quality as VMAT+5, while making both planning and delivery faster.
In this study, our in-house developed algorithm for automated multicriterial plan generation with integrated beam angle optimization (BAO) was used for NPC to explore development and assess added value of VMAT CSs with few fixed, non-coplanar arcs supplementing regular coplanar VMAT. VMAT+5 plans were used for selection of non-coplanar arcs. All generated plans are clinically deliverable on regular linacs.

Global study design
First, our in-house automated treatment planning system (TPS) was configured for NPC, in line with our clinical NPC planning protocol. Then the configuration was validated by comparison of clinically applied coplanar VMAT plans of previously treated patients (ClinVMAT) with corresponding automatically generated coplanar VMAT plans (AutoVMAT), to ensure use of high-quality automatically generated plans in this study. Next, the system was used to automatically generate VMAT+5 plans for the included patients, each consisting of coplanar VMAT supplemented with 5 non-coplanar IMRT beams with optimized patient-specific beam directions. Based on the overall distribution of selected patient-specific beam angles for the VMAT+5 plans, we then defined a non-coplanar VMAT CS, aiming at a reduction in the number of couch rotations during treatment compared to VMAT+5 (less than four). A second CS was defined as coplanar VMAT with one additional partial non-coplanar VMAT arc at couch angle 90 • (for symmetry reasons, as targets can be left-sided, central and right-sided). Partial noncoplanar arcs were used for linac-patient collision avoidance. Plans generated with the two CSs were compared with coplanar VMAT plans. The Rating guidelines for treatment planning studies [13] were used for study design and writing of the paper.

Patients and clinical protocol
In this study we used contoured planning CT-scans of 20 NPC patients that were recently treated at our center, all fulfilling the clinically used planning constraints, the prescribed PTV coverage, and treated with coplanar VMAT.
All patients were planned with a simultaneous integrated boost scheme, prescribing 70 Gy to the primary tumor and pathological lymph nodes (PTV70), and 54.25 Gy to the elective nodal areas (PTV54.25), delivered in 35 fractions. PTV54.25 was created by uniform expansions of the primary CTVs and the elective lymph node CTVs by 0.5 cm (clipped at the patient surface by 0.5 cm). PTV70 consisted of the primary CTVs, each expanded with a 0.5 cm margin (clipped at the patient surface by 0.5 cm). Average PTV70 and PTV54.25 volumes were 349 cc (range: 86-812 cc) and 751 cc (range: 184-1281 cc), respectively. The two PTVs of all included patients are presented in Figure A1 of Electronic Supplement A.
Dosimetric aims of the clinically applied planning protocol are presented in Table 1. The aim was to deliver 95% of the prescribed dose to at least 98% of the PTVs (V95%>98%), while maintaining PTV70 V107% below 2 cc (V107%<2 cc). Hard constraints were set for maximum doses of serial OARs: spinal cord, brainstem, brain, optical nerves, chiasm, retina, and mandible. While respecting the hard constraints, doses in parotid glands, submandibular glands (SMGs), oral cavity, swallowing muscles (organs associated with xerostomia and dysphagia), larynx, esophagus, cochleas and lenses were minimized, with parotid glands, SMGs, oral cavity, and swallowing muscles having highest priority.
Automated plan generation was validated by comparing the AutoVMAT plans with their corresponding ClinVMAT plans.
The system has an option for integrated optimization of beam intensity profiles and (non-coplanar) beam angle directions [8,11,12,15,[26][27][28][29][30][31][32][33], which was used in this study. Erasmus-iCycle generates Pareto-optimal and clinically favorable plans by applying an appropriate treatment site specific configuration ('wish-list') [15,17,34], containing hard constraints that can never be violated, and objectives that are optimized in order of priority. Wish-lists are generated through an iterative process where the initial aim is to mimic the clinical plans' quality, respecting the same guidelines, constraints and trade-offs. In later iterations, the wish-list is improved maximally, with a drive to surpass clinical plan quality [14].
When beam angle optimization (BAO) is combined with beam profile optimization, a candidate beam set has to be defined. Implicitly, the candidate beam set also defines the desired type of treatment technique, i.e., coplanar or non-coplanar.
The wish-list was developed based on the clinical protocol, using in total five training patients in two steps: an initial wish-list was created by extensive tuning with 2 patients (not included in the study cohort, to avoid bias), which was then followed by minor fine-tuning using 3 other patients (included in the study cohort as plan quality changes with the fine-tuning were very small). The final wish-list was then used to generate automated treatment plans for all other study patients. For more information on wish-list development for this study, see Electronic Supplement A, section A2.
For BAO in VMAT+5 plan generation, the full non-coplanar search space consisted of 437 candidate beams, separated by 10 • , excluding beams that could result in collisions between the patient situated at the treatment couch and the gantry, as verified at the treatment unit.

Plan evaluations and comparisons
After confirming that adequate PTV70 dose coverage (V95 > 98%) was obtained, all plans were normalized to the same PTV70 dose coverage (V95% = 98%, the clinical goal), in order to facilitate fair comparisons of all generated plans.
Plan evaluations and comparisons were mainly based on the clinically used parameters. For the parallel organs, the mean dose (Dmean) was evaluated; for serial organs, the D0.03cc was evaluated, as a more robust replacement of the Dmax [36]. In addition to the clinically Fig. 1. Differences between ClinVMAT and AutoVMAT in healthy tissue plan parameters; positive values in case of an advantage for AutoVMAT. For plan parameters with an * the difference is statistically significant. All plans were normalized for equal PTV70 coverage. Green boxes show interquartile ranges (IQR), separated by blue lines as median. The whiskers extend to a maximum of 1.5 × IQR beyond the box, to each side [38]. Values outside this range are plotted individually as outliers (+, in red). Subscripts L and R: Left and Right. Muscle S, Muscle M, Muscle I: superior, middle, and inferior swallowing muscles, respectively. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Table 2 Comparison of dosimetric plan parameters for AutoVMAT and differences with VMAT+5, VMATCS60 and VMATCS90. Mean values, standard deviations (SD) and ranges refer to the 20 patients in the study. Data related to statistically significant differences (p < 0.05) are in bold. NS: non-significant. Parallel OARs: parotid glands, SMGs, oral cavity, swallowing muscles, larynx, esophagus, cochleas and lenses. Serial OARs: spinal cord, brainstem, brain, optical nerves, chiasm, retina, and mandible. applied parameters, the dose bath was evaluated through the patient volumes for V10Gy, V30Gy and V50Gy. Statistical analyses were performed using the two-sided Wilcoxon signed-rank tests. Differences with p-value < 0.05 were considered statistically significant.

Comparison of automatically and manually generated coplanar VMAT plans
Development of the applied wish-list for automated plan generation for NPC is described in section A2 of Electronic Supplement A. The final wish-list can be found in Table A1 of Electronic Supplement A.
All automatically generated plans rescaled to 98% coverage for PTV70 (section 2.5) fulfilled all clinical constraints. Differences between ClinVMAT and AutoVMAT for healthy tissues are summarized in Fig. 1 and Table B1 (Electronic Supplement B) and patient-specific comparisons in Figure B1.
Automatically generated VMAT plans had overall favourable plan parameters compared to ClinVMAT. For OARs related to xerostomia and dysphagia we observed statistically significant differences for right parotid Dmean (reduction by 2.2 ± 4.8 Gy), left parotid Dmean (reduction by 3.9 ± 5.9 Gy), left submandibular gland Dmean (reduction by 3.8 ± 6.2 Gy), oral cavity Dmean (reduction by 2.6 ± 4.1 Gy), superior, middle and inferior swallowing muscles Dmean (reductions by 2.1 ± 2.3 Gy, 4.5 ± 3.8 Gy and 8.2 ± 7.4 Gy). Moreover, dose reductions with AutoVMAT for the larynx, mandible, left optical nerve and chiasm were statistically significant. ClinVMAT only performed statistically significantly better for the spinal cord, but in the AutoVMAT plans the spinal cord D0.03cc was always below the constraint level. Figure B1 shows for each of the 20 study patients separately, differences in plan parameter values. Figs B1a and B1b show consistent advantages of AutoVMAT for parotids, SMGs, oral cavity and swallowing muscles for all study patients, while

Fig. 2.
Differences between AutoVMAT and VMAT+5 (yellow), VMATCS60 (purple) and VMATCS90 (blue) in healthy tissue plan parameters; positive in case of advantage for VMAT+5, VMATCS60 and VMATCS90. For plan parameters with an * the difference is statistically significant. All plans were normalized for equal PTV70 coverage. Yellow, purple, and blue boxes show IQR, separated by blue lines as median. The whiskers extend to a maximum of 1.5 × IQR beyond the box, to each side [38]. Values outside this range are plotted individually as outliers (+, in red). Subscripts L and R: Left and Right. Muscle S, Muscle M , Muscle I : superior, middle, and inferior swallowing muscles, respectively. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) the level of sparing is highly patient-specific.
The comparisons between AutoVMAT and ClinVMAT demonstrate high plan quality for the automated treatment planning and supports application of the wish-list in this study for development of noncoplanar VMAT CS for NPC.

Establishment of non-coplanar class solutions
All VMAT+5 plans resulted in sufficient PTV coverage and respected the clinically applied dose constraints. AutoVMAT and VMAT+5 plans are compared in Table 2 and Fig. 2. Compared to AutoVMAT, VMAT+5 improved doses in several clinically important structures at the cost of some losses that were considered clinically less relevant, e.g. because the related structures only had a D0.03cc constraint that had to be respected (which was always the case) with minor clinical interest in further dose reduction.
The angular distribution of the selected IMRT beams in the VMAT+5 plans of the 20 study patients is presented in Fig. 3. Based on this distribution, the class solution VMATCS60 was defined with partial arcs at couch angles − 60 • and 60 • (purple boxes in Fig. 3). Partial arcs were used to avoid collisions. To investigate plan quality with the smallest possible number of non-coplanar arcs (one) for keeping treatment times as low as possible, the VMATCS90 class solution was defined with one partial arc at couch angle 90 • (blue box in Fig. 3).

Comparison of VMATCS60, VMATCS90 and VMAT+5 with AutoVMAT
All plans included in the analyses were rescaled to 98% PTV70 coverage, after confirming they had adequate PTV coverage and fulfilled the clinical constraints. Data is presented in Table 2 and Fig. 2. In the non-coplanar VMATCS60, VMATCS90 and VMAT+5, doses in several high priority structures reduced compared to coplanar AutoVMAT, at the cost of some clinically less relevant losses, mainly related to structures with a D0.03cc constraint that were anyway respected and mostly with low parameter values. As shown in Table 3, compared to VMATCS90, VMATCS60 had lower parotids Dmean (-0.8 ± 1.4 Gy and − 1 ± 2.1 Gy) and spinal cord and brainstem D0.03cc, while D0.03cc for the right optical nerve and the chiasm were on average higher (all D0.03cc values with constraints). Larger differences between VMATCS60 and VMATCS90 were patient-specific. Figure B2 (in Electronic Supplement B) shows for each of the 20 study patients separately, differences in plan parameter values between AutoVMAT and VMATCS60. Figure B2a and B2b show consistent dose reductions for the higher priority OARs for VMATCS60. For the lower priority OARs (non-constrained), the level of sparing is patient and OARspecific. For esophagus, patients 9, 10 and 12 have dose reductions of at least 10 Gy, while the average is 4.4 ± 5 Gy. However, for patient 2 and 9, cochlea doses were higher with VMATCS60. Nevertheless, in none of the plans OAR doses surpassed the clinically applied goal. AutoVMAT resulted in lower low dose bath and VMATCS60 in lower high dose bath, for all patients ( Figure B2f).

Discussion
High quality non-coplanar VMAT class solutions can in principle circumvent the current lack of algorithms for individualized noncoplanar arc selection in most commercial treatment planning systems, or prevent complex, time-and workload intensive selection of patient-specific non-coplanar arcs with manual planning. Alternatively, patients can be treated with regular coplanar VMAT, which may result in suboptimal dose for some tumor sites or for individual patients. The aim of this study was to use automated planning to develop and evaluate non-coplanar VMAT class solution for nasopharyngeal cancer, with only few non-coplanar static arcs to keep treatment times clinically feasible. One CS, consisting of a full coplanar arc supplemented with 2 fixed noncoplanar partial arcs at couch angles − 60 • and 60 • (VMATCS60) was derived from a population distribution of patient-specific IMRT beam angles obtained with our in-house developed algorithm for automated multi-criterial plan generation with integrated beam angle optimization. For comparison, the simplest non-coplanar CS with only one noncoplanar arc at couch 90 • was investigated as well.
To the best of our knowledge, the use of automated planning for development and evaluation of VMAT CSs has never been investigated for NPC. Many studies have demonstrated superior plan quality for automated planning compared to manual planning, for many tumor sites [14,[16][17][18][19][20][21][22][23][24][25][26][27]. Especially for development of CSs, high plan quality is required, as the aim is to in principle treat all future patients with the developed CS. A suboptimal CS would result in an on average suboptimal dose in these patients. The CSs investigated in this study and all plans generated for evaluation were created for regular linac delivery.
Prior to the investigations on CSs, our optimizer for automated plan generation was configured for NPC treatment in line with our clinical planning aims by creating a 'wish-list'. Automatically generated coplanar AutoVMAT plans were then compared to the clinically delivered coplanar VMAT plans (ClinVMAT). This first analysis pointed at an important opportunity for improving plan quality by replacing manual planning with automated planning, especially for structures related to xerostomia and dysphagia. Sparing of these structures was feasible while maintaining adequate target coverage and strictly obeying hard constraints on other OARs. This observed gain with autoplanning without any manual fine-tuning is remarkable, given the large anatomy variations in this patient group ( Figure A1).
For establishment VMATCS60, our optimizer was used to first automatically generate coplanar VMAT plans with 5 additional optimized patient-specific non-coplanar IMRT beams. The idea behind this approach was to identify the most important non-coplanar directions for NPC and to explore how they could be combined in few arcs. Although there were substantial variations in the patient group, there were two clusters of directions that pointed at an opportunity for VMATCS60 to improve plan quality. In a recent study, the applied consecutive beam angle selection in Erasmus-iCycle was compared to a new approach that uses so-called total-beam-space plans, confirming the high quality of automated beam angle selection in Erasmus-iCycle [33].
Overall, all investigated non-coplanar approaches i.e., VMAT+5, VMATCS60 and VMATCS90, outperformed coplanar AutoVMAT with respect to sparing of structures related to xerostomia and dysphagia (with the highest clinical priority), at the price of some losses that were considered clinically less relevant, e.g., because the related structures only had a D0.03cc constraint that had to be respected (which was always achieved) with minor clinical interest in further dose reduction.  Table 2 suggests (no p-values provided for direct pairwise comparisons) that dose reductions with VMATCS60 and VMATCS90 in xerostomia and dysphagia related OARs relative to AutoVMAT were larger than reductions with VMAT+5 relative to AutoVMAT, while for some clinically considered less important OARs, VMAT+5 seemed favorable. With this observation we conclude that there are no indications that VMAT+5 is dosimetrically superior to VMATCS60 or VMATCS90. As VMAT+5 also has 5 couch rotations per fraction instead of 2 (VMATCS60) or 1 (VMATCS90), the latter two approaches seem overall favourable, especially as they also avoid time-consuming patient-specific BAO. As shown in Table 3, VMATCS60 seems to have on average slightly better dosimetry than VMATCS90, but this advantage must be weighed against the use of 2 non-coplanar arcs instead of one. Figure B2 shows that the gain of non-coplanar treatment is highly patient-specific. This observation points at an option to generate for each patient in the treatment planning phase three plans: coplanar VMAT, non-coplanar VMATCS60 and VMATCS90 and then only select a non-coplanar approach in case of clinically relevant dosimetric advantages. With automated planning, generation of multiple plans per patient can be performed without excessive increase in planning time and workload [19].
The observed superiority of non-coplanar set-ups for NPC treatment confirms the findings by Wild et al. [4]. These authors applied different non-coplanar techniques to three nasopharyngeal tumour cases, including comparisons between coplanar VMAT, couch optimized noncoplanar VMAT and VMAT with dynamic couch rotation. While their results indicate that couch tilts do not yield as high plan quality as simultaneous couch and gantry rotations, the technology for rotating beam and couch trajectories is not currently commercially available.
Two investigators (JL, RB) independently filled out the RATING score list [13] for quality assessment of treatment planning studies, arriving at scores of 87% and 86% (maximum 100%).

Conclusions
In this study we used an in-house algorithm for automated treatment planning with integrated beam angle optimization to develop and evaluate non-coplanar VMAT class solutions for nasopharyngeal carcinoma, consisting of coplanar VMAT supplemented with only few fixed non-coplanar VMAT arcs to keep treatment times clinically feasible. Proposed non-coplanar class solutions with one or two non-coplanar arcs better spared xerostomia and dysphagia related OARs than coplanar VMAT. Also, the use of automated planning instead of manual planning resulted in dosimetric gains. Advantages of using a noncoplanar class solution and of using automated planning were both highly patient-specific. [37].

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.