The impact of a high‐definition multileaf collimator for spine SBRT

Abstract Purpose Advanced radiotherapy delivery systems designed for high‐dose, high‐precision treatments often come equipped with high‐definition multi‐leaf collimators (HD‐MLC) aimed at more finely shaping radiation dose to the target. In this work, we study the effect of a high definition MLC on spine stereotactic body radiation therapy (SBRT) treatment plan quality and plan deliverability. Methods and Materials Seventeen spine SBRT cases were planned with VMAT using a standard definition MLC (M120), HD‐MLC, and HD‐MLC with an added objective to reduce monitor units (MU). M120 plans were converted into plans deliverable on an HD‐MLC using in‐house software. Plan quality and plan deliverability as measured by portal dosimetry were compared among the three types of plans. Results Only minor differences were noted in plan quality between the M120 and HD‐MLC plans. Plans generated with the HD‐MLC tended to have better spinal cord sparing (3% reduction in maximum cord dose). HD‐MLC plans on average had 12% more MU and 55% greater modulation complexity as defined by an in‐house metric. HD‐MLC plans also had significantly degraded deliverability. Of the VMAT arcs measured, 94% had lower gamma passing metrics when using the HD‐MLC. Conclusion Modest improvements in plan quality were noted when switching from M120 to HD‐MLC at the expense of significantly less accurate deliverability in some cases.

For inversely optimized treatment plans, the added resolution of the HD-MLC theoretically increases the ability of the optimizer to produce a highly conformal treatment plan, sparing organs at risk (OARs) while providing adequate target coverage. However, this potential increase in plan quality depends on the ability of the optimizer to efficiently and effectively search the allowed solution space and additionally is fundamentally limited by the physical properties of the photon beam and MLC leaves.
Our department began using Varian's EDGE linear accelerator for stereotactic treatments in January of 2016. Two of the first treatment types to be transferred almost exclusively to the EDGE were spinal stereotactic body radiation therapy (SBRT) and spinal stereotactic radiosurgery (SRS), both delivered with volumetric modulated arc therapy (VMAT). The geometric complexity of these cases, often with irregularly shaped target volumes that may include multiple vertebral bodies and can circumscribe the spinal cord, makes them excellent candidates for treatment on this specialized accelerator. 2 With the transfer of these patients to the EDGE, a marked increase in plan modulation was observed, visualized both in the MLC-leaf trajectory sequence in the Eclipse treatment planning system (TPS) as well as quantified by the number of monitor units (MU) and our in-house VMAT complexity metric. 3,4 We also observed a sudden increase in plans failing patient-specific quality assurance with the ArcCHECK (Sun Nuclear Corporation, Melbourne, FL, USA).
Because there is often very little extra time built into the workflow for hypofractionated spine treatments, a failing pretreatment quality assurance measurement can lead to significant disruption in the planned treatment course. The case may either be replanned, which often leads to a delay (or multiple delays) in the patient start date, or the treatment team may elect to proceed with the initial treatment plan and accept the added degree of delivery uncertainty.
The goal of this work was to evaluate the change in plan quality and delivery accuracy when switching from the M120 to HD-MLC, and to test the effect of an added optimization objective aimed at reducing plan modulation on these quality indices.

| ME TH ODS
For this study, 17 previously treated spine SBRT cases employing the M120 were selected for study. A variety of sites (3 C-spine, 11 T-spine, and 3 L-spine) were chosen with geometries representative of that typically seen in our clinic. To ensure consistency in planning and plan quality, all cases chosen were replanned by an expert dosimetrist using the current clinical version of our treatment planning system, Eclipse version 13.6. Each of the cases was inverse planned with VMAT with the Photon Optimizer on a fine (1.25 mm) grid using 2 arcs per our institutional standard for spine SBRT. The final dose calculation was done on a 1 mm grid with Analytical Anisotropic Algorithm (AAA) version 13.6.23.
All plans had a prescription dose of 30 Gy delivered in three fractions and used 6 MV-only beams. All plans were reviewed by a physicist and physician with experience in spine SBRT. Once an acceptable plan was generated (henceforth referred to as the M120 plan), for each case, three additional versions of the plan were created, as shown in Table 1.
Consistent geometry was used across all plans within a given case (i.e., identical arc rotation length, collimator angles and jaw positions).
The M120(HD) plan was created by converting the M120 plan to HD-MLC by grouping pairs of leaves to identically match the original MLC pattern, and then adjusting the MLC positions outwards (making larger apertures) by 0.7 mm to account for the fact that the dosimetric leaf gap (DLG) for the HD-MLC is smaller compared to the M120. 5 The adjustment amount was chosen empirically by comparing the doses in the M120 plan to the M120(HD) plan. Edits to the MLC positions were performed using an in-house Matlab script (Mathworks, Natick, MA, USA).
The HD plan was created by reoptimizing the M120 plan with the HD-MLC, using identical optimization objectives. The HD(MU) plan was the same as the HD plan but had an added MU objective (a penalty on an MU value above a given threshold) to reduce the MU down to that of the M120 plan.
To evaluate plan quality and modulation complexity, the M120 plan was compared to the HD and HD(MU) plans. To evaluate delivery accuracy, we compared the M120(HD) plan to the HD and HD(MU) plans. This approach was taken so that in terms of deliverability, all plans could be compared on an equal footing: using the same beam model, calculation algorithm, and delivery and measurement system. This excluded any potential differences in the accuracy of each workflow step for the M120 MLC versus the HD-MLC from biasing our results. Additionally, with this approach, small deviations in the dose distribution due to changing MLCs (i.e., converting M120 to M120(HD)) had no impact on the results of the study.
To quantify plan quality, several metrics for the target and nearby OARs were used. Conformity index (CI) and gradient index (GI) were used to quantify dose coverage and falloff between the three types of plans. The conformity index used here was the Paddick index, given by: where TV PI is the target volume encompassed by the prescription isodose surface, PIV is the prescription isodose surface volume, and TV is the target volume.
The gradient index was defined as where PI 50% is the volume encompassed by the 50% dose isodose surface. 6 T A B L E 1 Description of plans used in this study.

M120
Original plan using M120 MLC M120(HD) M120 plan recreated with HD-MLC by pairing the MLC leaves HD M120 plan reoptimized with identical objectives on HD-MLC HD(MU) HD plan reoptimized with MU objective to reduce MU to that of the M120 plan To quantify plan complexity, we used a complexity metric developed in-house that analyzes an MU-weighted average of the leafside perimeter divided by the aperture area. 3 This metric was designed to quantify plan modulation independent of target size, plan dose, arc length, and MLC type, and is defined as, In all cases shown below, statistical significance in the comparison between plans was determined using the two-tailed Student's t-test. and HD(MU)). The cases are sorted by the MU of the M120 plan. One case (case 12) did not have an HD(MU) plan because the MU of the HD plan was already lower than that of the M120 plan.

| RESULTS
In the majority of cases, the MU-objective in the optimizer was able to reduce the MU in the HD(MU) plan down to that of the original M120 plan. The maximum value that can be used in the MU objective is 4000.
There were three cases where the M120 plan had MU >4000, which resulted in the HD(MU) plan having fewer MU than the M120 plan (see the left-most three cases in Fig. 1). The modulation complexity of the HD plan was consistently higher than the corresponding M120 plan.
Using the MU objective slightly reduced the modulation complexity in all but one case (case 3), but not to the level of the M120 plan.  Some of the comparisons are statistically significantly different, but the absolute difference may not be clinically relevant. The average gradient index in the HD plans was 0.19 lower than in the M120 plans. Figure 3 illustrates the potential difference in dose distributions for this magnitude of change in GI. There is no clear best type of plan according to these metrics.  Average global normalization passing rates for the 51 measured plans are also included in Table 4 for reference.

4.B | Deliverability
Portal dosimetry measurements show that the spine plans optimized with the HD-MLC have significantly degraded deliverability. Using global normalization (last row of Table 4), the passing rates of the plans generated for this study were very similar to the passing rates we observed with clinical plans prior to the initiation of this study.
Using local normalization, over two-thirds of the measured HD plans fell below a 95% pass rate threshold. The deliverability was somewhat improved with the added MU objective, but did not match that of the original M120 plan. This was expected as though the MU were lower for the HD(MU) plans, the modulation complexity scores were still high compared to the M120 plans. We note that all plans F I G . 4. Example portal dosimetry measurement for one case with arcs named CW (clockwise) and CCW (counter clockwise). The predicted images are shown on the left (for the M120 plan), with gamma maps (4%/1 mm, local normalization) on the right. The 10% dose threshold is outlined in red. Orange pixels indicate failing measurement points (gamma >1). Arc Number F I G . 5. Difference in per arc gamma passing percentages between the M120(HD) plans and the HD and HD(MU) plans. Positive values mean that the M120 (HD) arc had a higher passing rate than the comparison arc. The graph is ordered from highest difference between M120(HD) and HD to lowest difference between M120 (HD) and HD passing percentage.
T A B L E 4 Summary of portal dosimetry results. A passing arc is defined as a defined as an arc with a gamma pass rate of greater than 95%. A passing plan is defined as a plan where both arcs pass. All measurements are locally normalized unless otherwise stated.

M120(HD) HD HD(MU)
Average arc pass rate 99. | 101 used in the comparison of deliverability were calculated using the same beam model and delivered using the same delivery platform.
However, because the HD and HD(MU) plans are more modulated, they are likely to be more sensitive to any weaknesses in the beam model or MLC parameters such as transmission and DLG.

4.C | Comparison to previous studies
Many other groups have performed planning studies to examine the benefit of the smaller HD-MLC leaf width. Chae et al found that for artificially contoured spine lesions, the HD-MLC significantly improved GI but not CI, but that the results more notable for IMRT vs VMAT.
The biggest gains were seen for complex target shapes. 7 Tanyi et al performed a planning study comparing the M120 to HD-MLC for liver and lung with 3D, IMRT, and VMAT treatment plans and found small differences between the MLC types with the biggest difference being the faster dose falloff with the HD-MLC. 8  show that for a 2.5 cc radiosurgery volume treated with IMRT, CI and dose falloff were both improved with the smaller MLC.. 10 Kairn et al did not use the HD-MLC but studied delivery accuracy for the treatment of spinal metastases using portal dosimetry and radiochromic film. 11 They found that IMRT plans tended to be more deliverable than dosimetrically similar VMAT plans even though the delivery time was longer. Similar to the present study, the authors were able to improve delivery accuracy by limiting the MU of the plans.
We note that one weakness of this study is that the PDIP algorithm is a standalone algorithm that is configured independently of the AAA algorithm (although it does have some shared beam data), and therefore does not directly verify the dose calculation of AAA. 12 While our clinical ArcCHECK experience appears to support the findings in this work that deliverability is reduced with the HD-MLC, the ArcCHECK is a coarse measurement compared to the target size of most of the spine plans studied. Further analysis of these plans could be performed using film to more thoroughly verify the AAA dose calculation strengths and weaknesses.
The solution found by the optimizer for the M120 plans is part of the solution set available for the HD and HD-MU plans, however, the optimizer tends to choose more highly modulated plans. This is partially a result of the fact that the optimizer does not have direct feedback about the modulation in the plan, outside of the number of MU, and therefore is not able to directly limit the modulation. 3 The MU objective is not a sufficient tool for ensuring the deliverability of plans while at the same time effectively searching for an acceptable dose distribution.
In practice when using the HD-MLC, the user will not typically have a comparison M120 plan, so it is difficult or impossible to effectively judge the modulation complexity versus the plan quality to find an acceptable solution. Using a complexity metric like the one employed here aids in comparing newly created plans to past plans that have previously been measured for delivery accuracy.
Besides analyzing complexity postoptimization, planners need improved tools for limiting modulation when it is not needed in order to ensure an accurate delivery of the dose. One potential solution is to penalize complexity during the optimization in order to help guide the optimizer into the solution space that represents an acceptable compromise between plan quality and deliverability. 3 Though in this work the HD-MLC plans exhibit modest gains in the quality of the dose distribution, this improvement will come at a cost in terms of calculation accuracy, delivery accuracy, and delivery efficiency. Additionally, if calculation and delivery accuracy are sacrificed, one cannot be sure that the promised dosimetric gains are in fact realized.

| CONCLUSION
The HD-MLC theoretically allows improved dose distributions through added degrees of optimization freedom, however, care must be taken during the optimization process to avoid needlessly increasing the level of plan modulation. We have shown that increased plan complexity may lead to increased delivery uncertainty while at the same only modestly improving the dosimetric quality of the plan.