SpaceOAR© hydrogel rectal dose reduction prediction model: a decision support tool

Abstract Prostate cancer external beam radiation therapy can result in toxicity due to organ at risk (OAR) dose, potentially impairing quality of life. A polyethylene glycol‐based spacer, SpaceOAR© hydrogel (SOH), implanted between prostate gland and rectum may significantly reduce dose received by the rectum and hence risk of rectal toxicity. SOH implant is not equally effective in all patients. Determining patients in which the implant will offer most benefit, in terms of rectal dose reduction, allows for effective management of SOH resources. Several factors have been shown to be correlated with reduction in rectal dose including distance between rectum and planning treatment volume (PTV), volume of rectum in the PTV, and change in rectum volume pre‐ to post‐SOH. Several of these factors along with other pre‐SOH CT metrics were able to predict reduction in rectal dose associated with SOH implant. Rectal V55Gy metric, was selected as the dose level of interest in the context of 60 Gy in 20 fraction treatment plans. Models were produced to predict change in RV55Gy and pre‐SOH hydrogel RV55Gy. These models offered R‐squared between 0.81 and 0.88 with statistical significance in each model. Applying an ω1 = 3% lower limit of pre‐SOH RV55 Gy and an ω2 = 3.5% lower limit on change in RV55 Gy, retained 60% of patients experiencing the largest rectal dose reduction from the hydrogel. This may offer a clinically useful tool in deciding which patients should receive SOH implant given limited resources. Predictive models, nomograms, and a workflow diagram were produced for clinical management of SOH implant.

resulted in improved quality of life for patients receiving prostate EBRT. [5][6][7][8][9][10] SpaceOAR© hydrogel has been shown to reduce the rectal dose using both intensity-modulated radiotherapy (IMRT) 5,7,10,11 and volumetric modulated arc therapy (VMAT) techniques. 8,12,13 However, rectal dose reduction due to SOH is not equal in all patients. 14 Patients with clinical risk factors such as the presence of hemorrhoids or previous abdominal surgery have been shown to receive large benefits from SOH. 15 Hutchinson et al. performed a cost effectiveness study to include patient cost due to loss of income in addition to costs incurred due to medical intervention of acute and late side effects. 16 Results indicated SOH incurred an additional $518 cost with 3D-conformal radiotherapy (3D-CRT) while reducing cost by $2640 when treating with high-dose stereotactic body radiotherapy (SBRT) or stereotactic ablative radiotherapy (SABR). Additional cost effectiveness models have cited a large range in cost differential of SOH implementation. 16,17 Ensuring optimal management of SOH will allow for the most effective use of resource.
Volume of high dose to the rectum, such as the relative volume receiving 70 Gy (RV70Gy) in 78 Gy RT prescriptions, has been correlated with increased risk of rectal toxicity. 18,19 In recent years hypofractionated regimens have become increasingly common given the results of randomized trials. 20,21 There is currently no widely accepted clinically defined equivalent of RV70 Gy for hypofractionation. However, by using the linear-quadratic model to estimate a biological equivalent dose, adjusted slightly to account for the proportion of the total prescribed dose represented by the V70 Gy, the V55 Gy can be generated as a V70Gy approximation. Certainly, rectal doses in the range of 54.25-55.2 Gy appear clinically relevant with regard to toxicity in the setting hypofractionated regimens using 3 Gy per fraction. 22,23 The rectal V55Gy, therefore is a reasonable hypofractionated approximation of the V70 Gy, applicable to 60 Gy in 20 fraction prescriptions. The RV55 Gy represents a dose level of interest with hypofractionated prostate radiotherapy prescribed to 60 Gy in 20 fractions.
Several indicators have been correlated with rectal dose-volume and have been used in predictive models. Change in rectal volume from pre-SOH to post-SOH plans has been shown to be proportional to the pre-to postchange in rectal dose. 7 The distance from planning treatment volume (PTV) to rectum has been used to predict the lowest achievable rectum dose in IMRT prostate cancer treatment. 24 Finally, the overlap between the expanded PTV and the rectum has been related to reduction in rectal dose. 24,25 The primary aims of this study were to retrospectively determine the pre-SOH CT metrics which were strongly correlated with a change in rectal dose from pre-to post-SOH treatment plans and to create linear models that can be used in nomograms to determine a priori patients that stand to benefit the most dosimetrically from SOH implant. These models would be useful in directing patient selection and serve as a guide as to the achievable rectal dose reduction for a given patient. Secondary goal was to minimize the number of contours required on pre-SOH CT for these prediction models in order to create simple and useful decision support tools.

| METHODS
Some of the methods described in this study are similar to the methods used in a previous paper. 26 However, different analysis has been applied to the data resulting in unique results and conclusions.

2.A | Planning data sets
Anonymized CT data set of 22 patients with SOH implant between the rectum and the prostate were selected for this institutional ethics board approved retrospective study. Patients on this study received a CT scan 30-60 min prior to the fiducial and SOH implant with a comfortably full bladder and empty rectum. Patients were instructed to void and then drink 750 ml of water within a 15 min window, 1 h prior and perform a micro-enema 2-3 h prior to the pre-SOH CT scan. SOH was then implanted between the rectum and prostate transperinneally via ultrasound guidance followed by implantation of three to four gold fiducial markers. A post-SOH CT scan and a pelvic MRI were taken 1 week postimplant with patients following the same bladder and rectum instructions. The MR images (T1 FSPGR and T2 proton weighted) were registered with the post-SOH CT scan focusing primarily on the prostate and were used to contour structures of interest, in particular the SOH. One patient was removed from the studied data set due to an empty bladder in the pre-SOH CT scan.

2.B | Structure sets
Contouring of the prostate and seminal vesicles was performed by three experienced genitourinary radiation oncologists. Organs at risk were contoured by experienced radiation therapists and subsequently reviewed and modified as needed by the radiation oncologists. All contours were peer reviewed by the entire BC Cancer -Victoria genitourinary radiation oncology group for quality assurance.
A standardized structure set was used for each patient with the CTV, rectum, bladder, SOH, penile bulb, and femoral heads being contoured. CTV was contoured as the whole prostate gland and 1 cm of the seminal vesicles to emulate patients with intermediate risk disease. PTV was produced by adding margins of 7 mm in all directions with the exception of 5 mm in the posterior direction. The rectum was contoured from the ischial tuberosities to the rectosigmoid junction as a whole organ (rectum) and as a wall structure (3 mm inner wall structure created, RW). A second rectum wall structure was created for plan optimization as per the BC Cancer -Victoria planning guidelines (RW17.5). This is an adapted PROFIT 20 structure which stretches 17.5 mm superiorly and inferiorly of the most extreme PTV slices while maintaining a wall structure. Similarly, three bladder structures were produced: one as a whole organ (bladder), one as a wall structure (BW), and one following BC Cancer -Victoria planning guidelines (BW17.5). SOH and CTV structures were contoured on the post-SOH CT using the fused MR image as a guide. Penile bulb was contoured as the bulbous spongiosum below the GU diaphragm and proximal to the penile shaft. Femoral heads were contoured from the heads of the femur to the area between the greater and lesser trochanters.

2.C | Treatment planning
Treatment planning was completed using Eclipse version 13.6 to produce hypofractionated 60 Gy in 20 fraction single-arc VMAT plans for each CT dataset with final dose calculations being completed by the anisotropic analytical algorithm (AAA, version 11.0.31). Two VMAT treatment plans were created per patient by a single planner, one pre-SOH, and one post-SOH. Single-arc VMAT treatment plans were created using inverse optimization (progressive resolution optimizer version 11.0.31) with a series of optimization objectives on various structure contours. Rectal wall (RW) was used as an optimization structure as it has been shown to most effectively reduce the rectal dose when using SOH. 26 Similarly, optimization objectives were placed on PTV, CTV, bladder wall, penile bulb, and femoral heads. Treatment plans were deemed optimal once the plan evaluation objectives shown in Table 1 were reached with a plan normalization adjustment of less than AE0.5% following the final dose calculation. This strict limit on plan normalization adjustment compared to clinical practices (up to 5% adjustment for such hypofractionated prostate plans) was placed to reduce variation between plans due to plan normalization.
The plan evaluation objectives shown in Table 1 were taken from BC Cancer -Victoria planning procedures for hypofractionated, 60 Gy in 20 fractions, prostate radiotherapy. These goals were adopted from the PROFIT 20 and CHHiP 21 studies on hypofractionated prostate radiotherapy. The RW17.5 and BW17.5 structures refer to the secondary wall structure contoured in which both the RW and BW are limited to 17.5 mm superiorly and inferiorly of the final PTV slice.

2.D | Metrics of Interest
Several pre-SOH CT geometric descriptors were used as independent metrics in the development of multiple linear regression predictive models. Indepedent metrics used in the models were extracted from Eclipse TPS or RadOnc R package version 1.1.5 using RS and RD dicom files. Rectum, RW, PTV, and CTV volumes were all extracted from Eclipse TPS pre-SOH plan DVH text files. An additional structure was created by a boolean operation between the PTV and the rectum resulting in the volume of rectum in the PTV (RinPTV) in Eclipse TPS prior to file export. RinPTV and CTV volumes were also normalized by dividing their volumes by the rectal volume allowing for a relative metric as opposed to an absolute metric. RadOnc R package was a useful tool in accessing further geometric information from structures. 27 The distance between rectum and CTV was a structure of interest, as such the Hausdorff distance was measured. Hausdorff distance measures the furthest point between two 2D contours. 28 Hausdorff distance was determined along each CT slice to produce a mean distance between rectum wall and CTV (RWtoCTV). For this measurement, the height limited rectum wall, RW17.5, was used to ensure consistency in superior/ inferior direction.

2.E | Statistical analysis
Modeling began by calculating Pearson correlation coefficients to identify anatomical features well correlated to dependent metrics.
Independent metrics with high Pearson correlation coefficients were T A B L E 1 Plan evaluation objectives used in VMAT and IMRT treatment planning.

Structure
Metric (cGy) Volume IMRT, intensity-modulated radiotherapy; VMAT, volumetric modulated arc therapy. included in multiple linear regression models. Prior to modeling process, any independent metrics indicating dependence were identified and only one of them was selected for modeling purposes. As expected both PTV and CTV as well as rectum and rectum wall volumes were found to be dependent due to contouring methods. CTV volume was used over PTV volume as CTV represents a clinical region of interest while PTV is a geometric volume. Furthermore, contour variations are amplified when a PTV margin is applied. Rectum volume and rectum wall volume were used interchangeably, but separately.
All independent metrics were initially included in the multiple linear regression model of a selected dependent metric and then reduced through use of Akaike information criterion (AIC) and variance inflation factor (VIF). AIC represents the relative quality of the linear regression models with lower values indicating more effective models. A stepwise AIC approach was used in removing independent variables to generate an optimal linear regression model for each dependent metric (ΔRV55 and Pre-RV55). VIF quantifies the severity of multicolinearity between variables. Independent metrics with large VIF were also removed from the model until VIF was acceptable for all variables (VIF < 5). 29 Independent metrics chosen for multiple linear regression were associated with a beta coefficient, β. These coefficients combined with the metric creates a linear representation of the data as shown below.
In this expression, y is the dependent metric, x i are the independent metrics used in multiple linear regression models, and β i are the beta coefficients. The relative contributions of each of these independent metrics in describing variance of the data may be described through a structure coefficient, r s . 30 The structure coefficient describes the contribution of independent metrics to the variance of the dependent metric.
The structure coefficient is described by dividing Pearson correlation coefficient, r pears , by the square root of R-squared from the model. The squared structure coefficient, r 2 s , will be reported in this study.
Two models were produced per independent predictive metric: an advanced model with metrics defined using RadOnc package (Model 1) and a simpler model using metrics available in Eclipse TPS (Model 2). The selected models were evaluated using leave-one-out cross-validation (LOOCV) techniques reporting the predicted R-  Table 2 shows the measured min, max, mean, and standard deviation of the dependent rectal dose metrics for the 21 patients used to generate predictive models while Table 3 summarizes the statistics for the independent pre-SOH CT scan metrics. Table 3

3.A | Change in RV55Gy (Δ RV55) Models
Change in RV55Gy (ΔRV55), a relative dose-volume metric, was measured to be 4.5 ± 3.0 % with a range between 0.8% and 10.8% for the 21 patients selected demonstrating large variations in rectal dose sparing using SOH. ΔRV55 was predicted using normalized RinPTV, rectum volume, and RWtoCTV for Model 1 (P < 0.0001, R 2 =0.83) and with only normalized RinPTV volume for Model 2 (P < 0.0001, R 2 =0.81). Model coefficients are summarized in Table 4 with Fig. 2 presenting the predicted plotted against the measured values. Normalization of RinPTV volume produced higher Pearson correlation coefficients compared to RinPTV volume metric resulting in a metric with structure coefficients of about 1.00. ΔRV55 on its own does not offer enough information to make clinical decisions and must be combined with pre-SOH dose information.

3.B | Pre-SOH RV55Gy (Pre-RV55) Models
Pre-SOH RV55 Gy (Pre-RV55) models may offer insight to relative change in rectal dose, a metric often cited in SOH papers. Pre-RV55 was measured to be 5.9 AE 3.7% on average and ranged from 0.8% to 13.5% for the 21 patients included in the modeling. Pre-RV55 was predicted with the use of many metrics, including rectum or RW, CTV, and RinPTV volumes ( Table 5). The distance between RW and CTV was cubed and produced an effective correlation for Model 1 (P < 0.0001, R 2 = 0.87) but was omitted for Model 2 which only includes RW and RinPTV volumes as predictors (P < 0.0001, R 2 = 0.80) as summarized in Table 7.

3.C | Leave-One-Out Cross-Validation (LOOCV)
Results from LOOCV analysis resulted in predicted R-squared, MAE and %MAE values for each model as reported in the change in rectum volume from pre-to post-SOH plans was proportional to the change in rectal dose. In this retrospective study, rectal volume metrics such as rectum or RW volumes were also correlated with the change in rectal dose metrics ( Table 2).
The stepwise approach of linear regression modeling using AIC and VIF for model selection retained a variety of independent variables for each dependent metric. When modeling, there is always a risk of overfitting a modeldescribing the random error in the data instead of the relationship between variableswhen too many independent variables are used in the model causing an inflated Rsquared value. 31 In an attempt to combat this effect, a maximum of three independent variables were used to describe the dependent metrics given the limited number (n = 21) of datapoints. Two models were produced for each dependent metric: Model 1 and Model 2. ΔRV55 was found to be highly correlated with independent metrics such as RinPTV, normalized RinPTV, and RWtoCTV. RWtoCTV  Similar to ΔRV55, Pre-RV55 was highly correlated with RinPTV and RWtoCTV. As shown in Table 7, both these metrics resulted in the highest structure coefficients for both Model 1 and Model 2.
Both Pre-RV55 Models 1 and 2 yielded larger R-squared values (0.88 and 0.87, respectively) compared to the ΔRV55 Model 1 and Model 2 (0.83 and 0.81, respectively). The predicted R-squared values from LOOCV for Pre-RV55 were found to be slightly larger than for ΔRV55 at 0.76 and 0.79, indicating that these models are also able to effectively predict new observations. The measured and predicted values are also presented in Fig. 2 and lie near the ideal trendline. However, points with a low Pre-RV55 value seem to be slightly overpredicted while values lying near the center of the distribution are well predicted. LOOCV resulted in a %MAE of 11.62% for Model 1 and 10.65% for Model 2 which lie near the values reported for Δ RV55. Due to the low %MAE and high predicted Rsquared value, both models can effectively predict Pre-RV55.
Producing clinically useful decision support tools for SOH management was a priority when developing these predictive models. In both ΔRV55 and Pre-RV55 models, Model 2 resulted in a lower % MAE and higher predicted R-squared as shown by Table 8 The flowchart based on the above prediction models and associated cutoff limits removed a total of eight patients from the implant pool. The above cutoff limits are based on our single institution results and adjustments to these limits will remove greater or fewer patients from the implant pool. Due to the imperfection of the prediction models, one patient was incorrectly kept in the implant pool during the first selection step but correctly removed during the second step, and one patient was incorrectly removed during the second selection stage. These selection errors may be observed in the paired (measured-predicted) patient data sets shown in Fig. 6.
Patients with which the predictive value lay above the cutoff while measured was below or vice versa were incorrectly selected. This is due to the imperfection of the models as quantified by the %MAE, and as such when a value near the limit is predicted, there may be errors in the patient selection process. Additionally, there is significant patient benefit to these prediction models. It is important to consider that a small reduction in rectal dose may not be significant enough to put a patient through the discomfort (i.e., mild pain) and potential complications (i.e., low risk of sepsis) of the procedure. These predictive models and proposed cutoff limits aid in this patient selection.
The models derived in this study may offer effective SOH management guidelines, however, their limitations must be considered.
The maximum ΔRV55 from pre-to post-SOH observed was approximately 10% (Table 3) and as such any value predicted above this value may have a large associated error. Limits could be applied to independent geometric variables ( technique (i.e., SABR with implanted transponders or robotic delivery) with a narrow posterior margin (i.e., 2-3mm). Additionally, models were built using dosimetric information from plans where a specific, but commonly utilized, PTV margin was applied. PTV margins depend heavily on the type of image-guidance applied. 35 (IRS) to help identify patients in which IRS was not beneficial. This study offered a decision support tool which provides cost effectiveness analysis along with toxicity prediction. Models presented here predicted change in rectal dose using nomograms after a single pre-SOH CT scan and contouring of rectum and CTV structures. Implementing these linear models may require an additional CT scan prior to SOH implant if one was not completed during diagnosis and staging. This additional step in patient process increases dose, requires further resources for contouring, CT scan scheduling and may be subject to error from bladder and bowel preparation methods.
Identifying patients that would benefit dosimetrically earlier in the process would further aid in decision-making for both doctors and patients. Applying these models to diagnostic CT scans would improve the workflow of SOH management described in this study.
Such a task may be performed with radiomics packages available in languages such as R and Python. The present models may potentially be applied to a diagnostic CT scan if the appropriate bowel and bladder preparation is performed prior to imaging.

| CONCLUSION S
Predictive models were created for change in RV55 Gy (ΔRV55) and pre-SOH RV55 Gy (Pre-RV55). All models reached statistical significance with ΔRV55 and Pre-RV55 models reaching R 2 and predicted R 2 values greater than 0.75 and MAE between 9.6-12.9%. Volumes of rectum in PTV (RinPTV), as well as CTV and rectum volumes offered highest correlation with dependent metrics. Distance between rectal wall and CTV contours offered high Pearson correlation but was not included in TPS-based models (i.e., Model 2). Applying a lower limit of ω 1 = 3% on the Pre-RV55 model and ω 2 = 3.5% on ΔRV55 models retained approximately 60% patients receiving sufficient rectal dose reduction from the SOH implant (12/21). The mean post-SOH RV55 Gy for the retained patents (1.8 ± 1.7%) was found not to be statistically different from the mean pre-SOH RV55 Gy of those not retained (3.0 ± 1.9%) in the implant process. Linear predictive models along with specific limits offer decision support tools for more effective SOH implant management and may aid in the planning process.

ACKNOWLEDG MENTS
The authors acknowledge Island Prostate Center for providing funding for SpaceOAR© hydrogel units. One author acknowledge Natural Sciences and Engineering Research Council of Canada (NSERC) for providing funding for this research.

CONFLI CT OF INTEREST
No conflict of interest. LOOCV, leave-one-out cross-validation; MAE, mean absolute error.