Incorporating parotid gland inhomogeneity into head‐and‐neck treatment optimization through the use of artificial base plans

Abstract Despite a great improvement in target volume dose conformality made possible in recent years by modulated therapies, xerostomia remains a common and severe side effect for head‐and‐neck radiotherapy patients. It is known that parotid glands exhibit a spatially varying dose response; however, the relative importance of subregions throughout the entire gland has yet to be incorporated into treatment plan optimization, with the current standard being to minimize the mean dose to whole parotid glands. The relative importance of regions within contralateral parotid glands has been recently quantified, creating an opportunity for the development of a method for including this data in plan optimization. We present a universal and straightforward approach for imposing varying sub‐parotid gland dose constraints during inverse treatment planning by using patient‐specific artificial base plans to penalize dose deposited in sensitive regions. In this work, the proposed method of optimization is demonstrated to reduce dose to regions of high relative importance throughout contralateral parotids and improve predictions for stimulated saliva output at 1‐year post‐radiotherapy. This method may also be applied to impose varying dose constraints to other organs‐at‐risk for which regional importance data exists.

whole-mean dose is effectual for preserving an OAR (organ-at-risk) which exhibits a pure parallel functional architecture, given that the spatial variance of dose within the OAR is unimportant. The parotid gland was once believed to exhibit a pure parallel architecture and hence a spatially homogenous dose response, 12 and to this day, the current standard of care for minimizing the risk of post-RT xerostomia incidence for head-and-neck patients is to constrain the wholegland mean dose. 13 However, recent preclinical studies have demonstrated radiosensitivity within parotid glands to be inhomogeneous, having a spatially varying dose response throughout. [14][15][16][17][18] In a recent study, a model including voxel dose data as well as patient demographic and clinical pathology features 14 found the superior-anterior portion of the parotid gland to be the most influential in predicting xerostomia recovery. Furthermore, it was found that patients who developed xerostomia had a much higher mean dose to the inferior portion of the parotid gland. 14 Another study 15 used a tenfold cross validation test to show that dose to the region of the parotid gland containing stem/progenitor cells around the first branching of the Stensen's duct, was more predictive of xerostomia at 1 yr than dose to any other subregion of the gland. The same study also showed that the spatial distribution of dose in rat parotid glands affected salivary function recovery after treatment. Dose to the cranial 50% of the gland resulted in more than a 50% loss in salivary output, as well as tissue degeneration throughout the entire gland.
Clark et al. 16 partitioned contralateral parotid glands (CPGs) for a single cohort of 332 patients into 2, 3, 4, 18, and 96 equal volume subsegments and derived the relative importance of each from mean dose regressors using random forests and conditional inference trees. The parotid gland with the lowest mean dose was defined as the CPG. Parotid gland structure sets and dose profiles were used to calculate the mean dose to various subsegments, and outcomes were described using stimulated saliva output at 1-yr post-RT and selfassessed xerostomia questionnaires. For 18 subsegments, the most important subsegment (caudal-anterior) had a relative importance of 3.85 times the expected result for a homogenous parotid gland. The least important subregion exhibited virtually no importance. 16 Clark et al.'s model was chosen to be used for implementing spatially varying dose constraints for multiple reasons. For one, it is the only available model that maps out relative importance values throughout the entire gland. Furthermore, it derives importance data within the original reference frame of the parotid gland without transforming to an alternate reference frame. Subsegmentation into 18 equal volume subsegments of CPGs creates a desirable size for varying constraints, as it is large enough for dose to be effectively steered, while small enough to account for the varying importance within the gland.
The spatial inhomogeneity of the dose response within the parotid gland, if incorporated into external beam RT treatment planning, could reduce the risk of xerostomia for head-and-neck patients.
Studies have concluded that incorporation of nonhomogeneous effects into treatment planning can lead to improved outcomes. 19,20 The purpose of this work is to demonstrate the feasibility of a simple technique for including sub-parotid gland importance data into RT treatment plan optimization using artificial base plans (BPs).
To demonstrate the technique, we used Clark et al.'s 16 intra-parotid gland importance data.

| MATERIALS AND METHOD
The RapidArc TM optimizer in Varian Eclipse TM (Varian Medical Systems, Inc.) is equipped with the ability to incorporate earlier radiotherapy deliveries into optimization. Pre-existing spatial dose distributions can be loaded directly in as BPs during optimization, and the standard optimization workflow proceeds otherwise unaltered. We made use of this feature to apply a spatially varying dose constraint to the parotid gland to preferentially spare regions of high relative importance from excessive dose during radiotherapy.  21 The average minimum distance between the primary PTV and the CPG was 3.3 cm. The relative importance of all 18 subsegments was determined using Clark et al.'s population-level importance data. 16 Subsegments were labeled in order of decreasing relative importance as S 1 ! S 18 , where S 1 is the subsegment of highest relative importance. A subsegmented CPG is shown labeled in Fig. 1.
DICOMautomaton was used to create artificial dose distributions (base plans) for each patient which adhered to the following formula: 1. Dose to all voxels located outside the CPG is zero.
2. Dose to regions of overlap between the CPG and target volumes is zero. This ensures that the prescription dose and tumor coverage will not be impacted.

3.
Within each subsegment of the CPG, the dose is uniform 4. Dose to the region of highest relative importance (caudal-anterior, Fig. 1 was D 0 , and the dose to other subregions was D 0 I, where I is a scaling factor proportional to the relative importance of the region compared to the most important subregion. Five different types of BPs were created for each patient. D 0 was set to 10Gy, 20Gy, and 30Gy to create distributions with a linear scale I, and these BPs were named BP 10 , BP 20 , and BP 30 , respectively. The values of D 0 were chosen as they span the range of typical mean parotid gland doses, and having multiple values allows us to empirically determine which type of base plan is most effective for applying constraints. In addition to the three BPs mentioned, a fourth was identical to BP 20 for subsegments S 1 ! S 5 , while the dose to all other subsegments (S 6 ! S 18 ) were zeroed; a fifth was assigned 50 Gy to subsegments S 1 ! S 5 , and 0 Gy to all other subsegments.
These two plans were named BP 20,5 and BP top5 .
In Varian Eclipse™, each patient had five placeholder plans created for the five artificially constructed BPs that were imported for use in External Beam Planning. As a control, VMAT plans were retroactively optimized while adhering to standard clinical head-andneck protocols (both parotids whole mean < 25 Gy or 1 parotid whole mean < 20 Gy). For each patient, two arcs with opposite direction 360°gantry rotations and a difference of 60°in collimator rotation (30°and 330°) were used. Plans were then reoptimized using each artificial BP. Loading the BPs into the optimizer does not by itself implement a spatially varying dose constraint throughout the CPG, as the standard parotid dose constraint is on the wholemean dose. Therefore, an additional upper bound dose constraint must be placed on the CPG. This constraint, combined with the BP dose, provides a spatially varying dose constraint which preferentially restricts dose to subsegments of high relative importance. The ideal constraint to set depends on the individual anatomy of the patient and was chosen to be between 0 and 15 Gy over the maximum dose in the current BP. In this manner, the constraint imposed on a given region of the CPG has varying strength, depending on the region's dose in the BP. All clinical dose constraints were met for all plans. These plans are named P 10 , P 20 , P 30 , P 20,5 , and P top5 , corresponding to the use of BP 10 , BP 20 , BP 30 , BP 20,5 , and BP top5 .
The control plan optimized without a BP is referred to as P 0 .
To maximize validity of a comparison between different plan types, it was paramount to minimize interplan bias and dose variability within structures other than the CPG. All plans for a given patient were optimized toward approximately the same V98 (percent volume receiving at least 98% of the prescription dose) in the closest PTV to the CPG. Dose constraints to other OARS for any given patient were optimized according to clinical guidelines and were independent of plan type, and PTV coverage was adjusted minimally.
Doses to all OARs without PTV overlap were kept below the standard clinical constraints. Each plan for a given patient had the same constraint on the whole-mean dose of the CPG, and plans using BPs  16 Here, the spatial distribution of importance is illustrated from (a) the anterior, and (b) the posterior. Subsegments are labeled according to their importance in (c).
was created, where D i is the dose to subsegment S i , Δ i is the maximum loss in salivary output predicted by infinite dose to S 1 independently, and D 50i , n i are parameters fit to the data, representing the dose predicting a decline in salivary output of 1 2 Δ i , and the steepness of the curve. The predicted response when only dose to S 1 is considered (Eq. (1) with only i ¼ 1) is shown in Fig. 2.

| RESULTS
Within the CPG, the mean dose to all subsegments is shown for each plan type in Table 1  In general, dose to parotid gland subsegments of high relative importance, which tends toward the caudal end of the gland, was reduced when planning with BPs as seen in Fig. 3. P 20 , P 30 , and P top5 significantly reduced dose to the F I G . 2. The predicted stimulated saliva output according to the Clark et al. 16 model at 1-yr post-RT relative to baseline is shown for subsegment S 1 .
T A B L E 1 The mean dose in each subsegment of the contralateral parotid gland for each plan type is shown. A subscript "s" represents a significant (P < 0.05) reduction in dose, while subscript "si" represents a significant increase in dose.   caudal-medial portion of glands were most prone to overlap. The frequency of overlapping for various subsegments is summarized in Table 3.

Subsegment
Optimizing with BPs did not prevent clinical dose constraints for

| DISCUSSION
The base plan approach for incorporating various intra-parotid gland dose constraints into head-and-neck RT plans through the additional constraint on the CPG. Furthermore, adding a unique dose constraint for each subsegments structure may have a noticeable impact on the optimization time.
We used population-level relative importance information of 18 equal volume parotid gland subsegments, 16  as a result, small increases in their already high doses had no effect on patient outcomes. Oral cavity mean doses were kept below clinical recommendations, so the statistically significant increase in mean dose seen is likely to be clinically insignificant.
The model by Clark et al. 16 has yet to be clinically validated since it was derived. We believe it to be a valid quantitative model as saliva predictions using segmentation into 18 subregions had values for mean-absolute-error and root-mean-square-error which are comparable to values when predictions were made using whole-mean dose. 16 Clark et al.'s model was used in our methods as it was the most favorable for designing constraints,however, the base plan method for imposing dose constraints can be applied using data from other models of regional importance for an organ at risk. The emphasis of this work is on the method of incorporating subregional dose constraints using artificial base plans that are specifically designed based on model data.
A challenge in this study was establishing a valid interplan comparison, as the varying initial dose conditions contained in each base plan ensures that optimized dose profiles for plans are nonidentical both inside and outside the CPGs. To minimize systematic errors, dose constraints to all OARs other than the CPG and all PTVs other than those in proximity to the CPG were set to the same value in different plan types, and the same V98 goal for bordering/overlapping PTVs was set in each case. Plans were created in a random order by a single planner for all patients. However, variability in the optimization process for each plan was impossible to eliminate entirely and could have impacted this study.
In the future, we wish to extend this work by imposing suborgan optimization criteria directly through the scripting API of treatment planning systems. The prospect of including dose constraints without requiring artificial base plans is appealing, but presents additional challenges as Varian Eclipse TM 's scripting API requires that automatic optimization proceeds inside a separate scripting window, meaning that dose constraints must be rigidly preset by the treatment planner. On the contrary, the approach presented here allows the treatment planner full and regular access to the standard optimization window, so that constraints can be adjusted throughout the course of optimization.