Improving accuracy for stereotactic body radiotherapy treatments of spinal metastases

Abstract Purpose Use of SBRT techniques is now a relatively common recourse for spinal metastases due to good local control rates and durable pain control. However, the technique has not yet reached maturity for gantry‐based systems, so work is still required in finding planning approaches that produce optimum conformity as well as delivery for the slew of treatment planning systems and treatment machines. Methods A set of 32 SBRT spine treatment plans based on four vertebral sites, varying in modality and number of control points, were created in Pinnacle. These plans were assessed according to complexity metrics and planning objectives as well as undergoing treatment delivery QA on an Elekta VersaHD through ion chamber measurement, ArcCheck, film‐dose map comparison and MLC log‐file reconstruction via PerFraction. Results All methods of QA demonstrated statistically significant agreement with each other (r = 0.63, P < 0.001). Plan complexity and delivery accuracy were found to be independent of MUs (r = 0.22, P > 0.05) but improved with the number of control points (r = 0.46, P < 0.03); with use of 90 control points producing the most complex and least accurate plans. The fraction of small apertures used in treatment had no impact on plan quality or accuracy (r = 0.29, P > 0.05) but rather more complexly modulated plans showed poorer results due to MLC leaf position inaccuracies. Plans utilizing 180 and 240 control points produced optimal plan coverage with similar complexity metrics to each other. However, plans with 240 control points demonstrated slightly better delivery accuracy, with fewer MLC leaf position discrepancies. Conclusion In contrast to other studies, MU had no effect on delivery accuracy, with the most impactful parameter at the disposal of the planner being the number of control points utilized.


| INTRODUCTION
Stereotactic body radiotherapy (SBRT) describes extracranial treatment techniques which utilize a larger delivery of radiation dose than conventional radiotherapy and in fewer fractions, resulting in a higher biological effective dose for the treatment site. 1 As with stereotactic radiosurgery for brain metastases, 2 the SBRT technique is able to provide extra dose to the target volume without exceeding recommended normal tissue tolerances. In the case of spinal metastases, the spinal cord may have already been irradiated through conventional radiotherapy and SBRT provides a noninvasive treatment option. 3 In order to limit dose to the spinal cord, modulated treatment techniques are used to produce complex dose distributions that spare organs at risk. [4][5][6] SBRT spine treatment can be delivered using a number of treatment machines like helical tomotherapy units, robotic radiosurgery systems or on a linac utilizing intensity-modulated radiotherapy (IMRT) or volumetric-modulated arc therapy (VMAT). In an Eclipse treatment planning system (TPS) & Varian linac environment, single arc VMAT has been found to produce inferior target coverage and normal tissue sparing compared to IMRT, but two arc VMAT is able to produce comparable results. 7 Comparisons between flattening filter free (FFF) and conventional modalities have demonstrated significant improvements in normal tissue/spinal cord sparing for FFF plans with a greater number of control points, while achieving the same level of target coverage. 8,9 While these and other planning studies have been conducted for SBRT spine, there are still gaps in the literature regarding the overall optimization of the SBRT spine technique, which requires further investigation.
A couple of planning studies have also included quality assurance (QA). A recent study found accuracy improvements for an individual SBRT spine case when the gantry spacing between control points was decreased from 4°to 3°. 10 The delivery accuracy of different IMRT and VMAT techniques has also been investigated for spinal treatment. 11 IMRT was measured to deliver more accurately than VMAT due to its reduced complexity. VMAT beams can be delivered more accurately if optimized with a monitor unit (MU) limit, which can produce a less complex plan. However, these findings may be unique to the dose calculation and optimization algorithms of the Eclipse TPS and the accuracy of the Varian linac setup. The study also allowed a slightly higher cord dose which could make large impacts in plan conformity. Another recent study investigated the effect of high definition MLCs on plan quality, complexity and deliverability 12 and found that while slight improvements to quality could be made for VMAT SBRT spine treatments, plan complexity increased and thus delivery accuracy suffered.
There is an overall lack of studies which combine both planning optimizations and QA results for SBRT spine treatment. This study aims to address this and improve upon previous studies through utilization of four different methods of QA. It is important to fully assess inaccuracies/accuracies in the TPS, so that it can be used confidently in the correct range of parameters.

| MATERIALS AND METHODS
Four previously treated and anonymized SBRT spine CT and contour datasets were selected with differing vertebra and target volume shapes (targeting had been previously performed by radiation oncologists according to international guidelines, 3,13 CT was performed with 2 mm slices). Eight different SBRT spine plans were created on each CT image dataset, utilizing both the 6 MV and 6 MV FFF beam.
Control points were set as planning constraints before inverse optimization and plans of 50, 90, 180, and 240 control points were created ( Table 1). The 50 control point plans consisted of nine beams delivering IMRT, while the 90, 180, and 240 control point plans were 1, 2, and 2 arc VMAT plans, respectively. Plan complexity was measured through different metrics: (a) modulation index (MI) which considers the fluence map (higher MI implies higher complexity), 14 (b) modulation complexity score (MCS) which considers aperture area variability and leaf sequence variability (higher MCS implies lower complexity), 15 and (c) the small aperture score (SAS) which considers the fraction of fields used <10 mm (higher SAS corresponds to more small fields). 16 The TPS utilized was Pinnacle3 ® 9.10 (Koninklijke Philips N.V., Amsterdam, The Netherlands). Plans were created within the TPS each with the same fractionation and dose constraints currently used clinically: 30 Gy in 3 fractions, with 100% of planned target volume (PTV) to be covered by 80% of prescribed dose, 17 but ideally covered by 90% as per RTOG 0631. 13 The other main metric concerned was assessment of the Paddick conformity index (CI) for target coverage (CI <1.0 implies poorer conformity) 18 as per ICRU 91. 19 Other departmental metrics concerned were PTV mean dose below 110% of prescription and 95% of the clinical target volume (CTV) to receive 98% of the prescription 1 (CTV-PTV margin of 2 mm). However, critical nervous structure (CNS) constraints always took priority. The main CNS structure concerned was the thecal sac with a maximum dose of 20 Gy. 20 Each of the 32 clinical treatment plans created was given the same iteration schedule for inverse planning. The dose calculations were made using the collapsed cone convolution (CCC) algorithm on a 2 mm dose grid 21 with a minimum leaf separation of 0.5 cm. The optimization schedule began with an initial 80 iterations after which the plan was assessed and the optimization objectives adjusted in order to achieve the desired conformity. After these adjustments,  22 ) and the ArcCheck were then compared to respective dose maps from the TPS through gamma analysis in SNC Patient with respective gamma criteria of 3%/1.5 mm and 2%/2 mm. All gamma analyses in this study were performed through absolute dose comparison with a local calculation and a 10% low dose cut-off threshold. 23 Finally, QA on the treatment plan delivery was also performed through the PerFraction ™ software package (Sun Nuclear Corp., Melbourne, FL, USA). PerFraction takes the MLC log-files from the linac and reconstructs the dose delivered on the CT image dataset using its own CCC style algorithm. A gamma analysis of the two calculated dose distributions was performed (1%/1 mm) for each plan. Correlation of data was assessed using the Pearson correlation coefficient, r, 24 given by where n is the number of samples and x i and y i are the samples.   Average planning computation times and treatment delivery times are shown in Fig. 3. Computation times scale according to the number of control points utilized. Plan delivery times were quicker for FFF due to the higher dose rate and increased with the number of arcs or beams used.
As an example, a plan that was delivered with a high level of accuracy was T11 240X with its film QA shown in Fig. 4, while a particularly poorly delivered plan, L2 90X, is shown in Fig. 5.
ArcCheck, Film, PerFraction, and mean pass rates plotted against control points are shown in Fig. 6. Film results would indicate that   MCS and MI had no significant trend with MU (r = 0.22, P > 0.05) and while the median IC result was a −0.05% difference to TPS, isocenter dose differences became more negative as MU increased (r = 0.46, P < 0.01). Correlation data have been collated in Table 2.

| DISCUSSION
Given the CNS prioritization, it is not surprising that very few plans achieve the target goals (Fig. 1). However, such plans are still clinically deliverable as 27 Gy in 3 fractions. Contrary to other studies, no difference was found in conformity between 6 MV and 6 MV FFF plans, 9  It can be seen that better quality plans only consistently resulted from increases to the number of control points (90 to 240), so it is of interest then that as control points increased, the likelihood of producing a plan with a greater SAS also increased [ Fig. 2(c)]. Without fully analyzing the impact of SAS on plan quality, one might be tempted to deduce that small fields produce better quality plansthis was found not to be the case (r = 0.15, P > 0.05). The most complex plans resulted from setting control points to 90, causing the VMAT optimization algorithm to produce more complicated MLC shapes in order to meet planning objectives. IMRT plans were not as complex, even though there were fewer control points, due to optimization algorithm differences and as such, the delivery accuracy was high.   11,12,16 and is most likely due to MLC leaf position discrepancies becoming more frequent 37 rather than model deficiencies.
As expected, plan complexity had no statistically significant impact on isocenter point dose measurements, as dose point placement followed the principles of ICRU 50 to ensure accuracy. 40 SAS had no effect on the gamma pass rates in contrast to the previous studies concerning IMRT plans in Eclipse. 16

| CONCLUSION
A planning and QA study of 32 SBRT spine plans, optimized on four vertebral sites, was conducted in order to find optimum conditions for plan conformity and accuracy of delivery. Plan complexity metrics were found to be independent of plan quality, while QA through four independent methods demonstrated that plan conformity was influenced by the number of control points.
In contrast to other studies, it was not MU that improved delivery accuracy but, rather, control points. Like studies on Varian/Eclipse systems, more complex plans were less accurate to deliver, which is attributed to MLC leaf position discrepancies, not solely small fields. The optimum SBRT spine plans in terms of both plan conformity and coverage as well as plan delivery accuracy came from increasing control points to 240.

CONF LICTS OF INTEREST
The authors have no other relevant conflicts of interest to disclose.