A Nomogram to predict parotid gland overdose in head and neck IMRT

To generate a nomogram to predict parotid gland (PG) overdose and to quantify the dosimetric benefit of weekly replanning based on its findings, in the context of intensity-modulated radiotherapy (IMRT) for locally-advanced head and neck carcinoma (LAHNC). Twenty LAHNC patients treated with radical IMRT underwent weekly computed tomography (CT) scans during IMRT. The cumulated PG dose was estimated by elastic registration. Early predictors of PG overdose (cumulated minus planned doses) were identified, enabling a nomogram to be generated from a linear regression model. Its performance was evaluated using a leave-one-out method. The benefit of weekly replanning was then estimated for the nomogram-identified PG overdose patients. Clinical target volume 70 (CTV70) and the mean PG dose calculated from the planning and first weekly CTs were early predictors of PG overdose, enabling a nomogram to be generated. A mean PG overdose of 2.5Gy was calculated for 16 patients, 14 identified by the nomogram. All patients with PG overdoses >1.5Gy were identified. Compared to the cumulated delivered dose, weekly replanning of these 14 targeted patients enabled a 3.3Gy decrease in the mean PG dose. Based on the planning and first week CTs, our nomogram allowed the identification of all patients with PG overdoses >2.5Gy to be identified, who then benefitted from a final 4Gy decrease in mean PG overdose by means of weekly replanning.


Introduction
During the course of intensity-modulated radiotherapy (IMRT) for head and neck cancer (HNC), large anatomical variations may result in delivered doses differing from the planned dose [1]. The literature shows that while dose variations in the clinical target volume appear extremely low [2][3][4][5], the percentage of patients with estimated PG overdoses ranges widely from 5 to 70 % [1,[5][6][7][8][9][10]. With the aim of correcting these PG overdoses, an adaptive radiotherapy (ART) strategy involving one or several replannings during treatment has been investigated [1,2]. These replannings are, however, time-consuming, as a complete delineation can take up to 2.5 h [11][12][13] and may not be beneficial for all patients. It is therefore crucial to identify patients with PG overdose and evaluate how ART benefits each individual. Ideally, replanning decisions should be based on early and simple anatomical criteria, such as weight loss or decrease in neck diameter, which have been identified as risk factors for overirradiation [10,[14][15][16][17]. However, a clear correlation between these markers and PG overdose has not yet been established [6]. After having identified early predictors of PG overdose, this dosimetric study had two objectives: 1) to generate a nomogram so as to predict PG overdose; 2) to quantify the benefits of weekly replanning, triggered by the nomogram, in terms of dose sparing and decrease in xerostomia risk.

Patients and tumors
A total of 20 patients (mean age: 63; range: 50-77) were enrolled in this study. All patients presented with locally-advanced oropharyngeal cancer (Stage III or IV, American Joint Committee on Cancer 7th ed.). Patient, tumor, and treatment characteristics have been provided in Table 1.

Treatment and planning
All patients underwent IMRT with total doses of 70Gy (2Gy/fraction/day, 35 fractions), combined with a simultaneous integrated boost technique [18] and concomitant chemotherapy (cetuximab or platinum). Planning CTs (CT0) were performed with intravenous contrast agents using 2-mm slice thickness, from the vertex to the carina. Three target volumes were generated: CTV 70, CTV 63 , and CTV 56 . The 70Gy clinical target volume (CTV 70 ) was equal to the gross tumor volume plus a 5-mm 3D margin, adjusted to exclude the air cavities and all bone mass free of tumor invasion. CTV 63 corresponded to the area at high-risk of microscopic spread, in particular the ipsilateral nodal level II, while CTV 56 corresponded to the low-risk subclinical area. The planning target volume (PTV) was the CTVs plus a 5-mm 3D margin, limited at 3 mm from the skin surface in order to avoid the build-up region and therefore limit skin toxicity [19]. The minimum PTV covered by the 95 % isodose line was 95 %. Dose constraints were set according to the GORTEC group (the French group of radiation oncology for head and neck cancer) ( Table 2).
Parotid sparing was not conducted if considered to the detriment of PTV coverage or other essential organs at risk (OARs).
During the treatment course, set-up errors >5 mm were corrected by weekly in-room stereoscopic kV imaging. Informed consent was obtained from all patients. This study was approved by the institutional review board (ARTIX study NCT01874587).

Weekly dose estimations
Each patient underwent six weekly CTs (CT1-CT6) using the same protocol as CT0 over the treatment course, except for some variations in intravenous contrast agent use, which was not systematically employed, particularly not in the context of cisplatin-based chemotherapy. Anatomical structures were manually segmented on each weekly CT by the same radiation oncologist for each patient. In the event of complete response, the original macroscopically-involved areas were still included in CTV 70 , which was adjusted to exclude any air cavities and bone mass showing no evidence of original tumor invasion. Treatment always commenced on Mondays, with each weekly CT performed the following Monday. As patients were treated 5 days per week, each weekly CT corresponded to a 10Gy additional dose to the PTV (CT1 at 10Gy, CT2 at 20Gy, and so on). The actual doses delivered weekly were estimated by calculating the dose distribution on the weekly CTs using treatment parameters and the CT0 isocenter ( Fig. 1).

Total cumulated dose estimations using deformable registration
For all patients, the weekly CT images were first registered to the planning CT using a rigid transformation defined by six parameters (three translations and three rotations). The mean squared error was used as a similarity criterion. The cumulated dose estimate relied on deformable image registration, using the free-form deformation (FFD) method. The control points were iteratively displaced according to the considered metric. The dense deformation field was obtained by B-spline interpolation [20]. The mutual information metric was used to handle the modified intensities between CT datasets caused by the presence of contrast agent. The geometric transformation obtained using both rigid and deformable registration was then applied to the weekly dose distributions in order to propagate each one to the planning CT0 dataset. The implementation was provided by the ElastiX library [21].
The average Dice score for PG registration was 0.81 (0.62-0.94). The propagated dose distributions were totaled to compute the cumulated dose on the planning CT, which was finally compared to the planned dose.
Linear regression model and nomogram to predict PG overdose PG overdose was calculated as the difference between the cumulated mean PG dose and the mean PG dose of the planning CT dataset (CT0).
The following anatomical and dosimetric parameters have previously been described in the literature as correlating with PG overdose [6,10,14,15], and were assessed for this correlation: CTV70 (cm3), PG volume (cm3), neck thickness (mm), PG-to-CTV70 distance, and mean PG dose. The PG-to-CTV70 distance was defined as the minimal Euclidean distance between the surfaces of the two contours. The difference between and ratio of each of these parameters computed from the two CTs (CT0 and each weekly CT, respectively) were also studied. The Pearson correlation was used to assess the correlation between variables significantly correlated with the PG overdose. When high correlation was observed between two variables (r 2 > 0.5), only the most significant parameter was included for further analysis. Finally, a linear regression method with backward elimination (coefficient of determination r 2 > 0.3, p <0.05) was used to generate a model for PG overdose prediction. Regression was run with and without an intercept. The standard errors were compared to decide whether ordinary least squares or regression through origin provides a superior fit [22]. The model's accuracy was then validated by the quantiles-quantiles plot (QQ-plot) and r 2 for PG overdose prediction. A nomogram was generated based on this model, i.e. a chart representing a linear function calculating a predicted value from plotted input data (Fig. 2).
Leave-one-out cross validation was then performed to estimate the model's stability and accuracy. This method consisted in all patients but one (n = 19) being used to develop a PG overdose model (difference between cumulated PG dose and planned dose). A PG overdose prediction was then calculated for the one remaining patient using the model. This predicted PG overdose was then compared to the cumulated PG overdose. This step was repeated for each patient. The variance of each model parameters was calculated to estimate the model's stability and identify outliers. The mean squared error of the predicted values was calculated.
Statistical analysis was carried out using the Statistical Package for the Social Sciences V. 20.0, and R language and environment for statistical computing.
Weekly replanning for the patients at risk of PG overdose Using the nomogram, we identified patients predicted to receive at least one PG overdose, for whom we performed a weekly IMRT replanning on each weekly CT dataset, in accordance with the dose constraints described for the initial planning. PTV coverage did not significantly differ between the initial planning and the weekly replanning. The dose constraints specified for the OARs complied with the GORTEC recommendations for all replanning, as for the initial planning. The mean PG cumulated doses with replanning were compared to those without. The impact of the replanning on the risk of xerostomia was estimated by using the Lyman Kutcher Burman (LKB) model of normal tissue complication probability  Step 1: The weekly delivered doses were calculated from the weekly computed tomography scans (CTs).
Step 2: A deformable image registration was applied to the weekly doses according to the deformation field between the weekly and planning CT (CT0).
Step 3: The propagated dose distributions were totaled to compute the cumulated dose for the CT0.

Results
A total of 37 PGs were analyzed, due to three ipsilateral PGs included within the PTV being excluded from analysis. Based on the difference between the cumulated and planned PG doses, two PG subgroups were identified (Fig. 3): a PG overdose group: 70 % of all the PGs, with a mean dose increase of 2.5Gy (up to 11.7Gy) and 16 patients presenting at least one overdosed PG; a PG under-dose group: involving the other 30 % of the PGs, with a mean dose decrease of 1.2Gy (up to 3.1Gy).
When identifying PG overdose predictors, we found the parameters from the first weekly CTs to be the most significant. Two anatomical and three dosimetric parameters were significantly correlated with PG overdose ( Table 3).
As the mean PG dose for CT1 and the CT1/planning mean PG dose ratio were highly correlated with the difference between the mean PG doses for CT1 and CT0 (ΔDosePG) (r 2 = 0.5 and 0.9, respectively, p <0.001), only the parameter with the highest r 2 was used (mean PG dose difference). The three parameters used for PG overdose prediction were CTV 70 on the planning CT dataset (CTV 70_CT0 ), the CTV 70 CT0-to-CT1 difference (ΔCTV 70 ), and the difference between the mean PG doses of CT1 and CT0 (ΔDosePG).
Based on these parameters, the resulting linear regression model for PG overdose prediction, optimized for all patients, was: Fig. 3 Parotid gland overdose assessment. The mean dose difference was calculated between the estimated cumulated dose (without replanning) and the planned dose, in each parotid gland (PG) (ipsilateral and contralateral), for each of the 20 patients. A dose difference with a positive or negative value corresponded to a PG overdose or under-dose, respectively. Predicted PG overdose calculated by the nomogram are represented by circles (ipsilateral PG) or diamonds (contralateral PG) (cf. Section 3.2) The corresponding nomogram is shown in Fig. 4a. The quantiles-quantiles plot of the nomogram is shown in Fig. 4b.
The correlation between the observed and predicted cumulated PG doses is shown in Fig. 5 (r 2 = 0.75).
Dose variations (PG over-or under-doses) were correctly predicted for 25 of the 37 PGs (Fig. 3). Of the 16 patients with at least one overdosed PG, 13 were accurately identified (One patient with an error concerning the side of the PG overdose). In the three patients (Patients 12, 18 and 19) who were inaccurately classified, the mean PG was increased by an average of 1Gy (range: 0.4-1.5Gy).
The model's performance, evaluated using the leaveone-out cross validation, in terms of identifying patients with or without PG overdose achieved sensitivity, specificity, and positive and negative predictive values of 80 %, 60 %, 86 %, and 50 %, respectively. The mean values and standard deviations (SDs) of each coefficient, considering the leave-one-out cross validation, were 0.007 (SD: 0.0006), −0.045 (SD: 0.004), and 0.507 (SD: 0.01), for CTV70 CT0 , ΔCT V 70 , and ΔDosePG , respectively. The mean square error for the predicted PG overdose was 2.6 Gy (SD: 1.6 Gy). No significant outliers were extracted from the validation procedure.
For the 14 patients identified by the nomogram as having a predicted PG overdose, the mean PG dose without ART was 34.8Gy (range: 20.9-51.4Gy), corresponding to a mean xerostomia risk of 37 % (20-86 %). The dosimetric benefit of weekly replanning for these 14 patients is shown in Fig. 6. Replanning achieved an average decrease of 3.9Gy in the mean PG dose (range: 0-9.5Gy), representing an average decrease in absolute xerostomia risk of 8 % (0-22 %).

Discussion
We generated a nomogram aimed at predicting PG overdose based on early predictors calculated on the planning CT dataset and on a CT performed in the first week of treatment.
In total, 14 of the 16 patients with a calculated mean PG overdose of 2.5Gy were identified by the nomogram. All patients with a PG overdose >2.5Gy were identified. Weekly replanning of these 14 targeted patients enabled the mean PG dose to be decreased by 3.9Gy compared to the cumulated delivered dose, corresponding to an 8 % decrease in the estimated absolute xerostomia risk.
Due to anatomical variations occurring during the course of IMRT, some PGs can receive doses exceeding the planned dose (overdose). In an attempt to prevent this, ART is designed to take these anatomical variations into account by generating one or several new plannings. An increasing number of studies demonstrate this technique to have dosimetric benefits [1,2,4,25]. The clinical impact of replanning has been evaluated in two studies [5,26], where it was shown to improve both patient quality of life [26] and localize disease control, yet had no impact on overall survival. Nevertheless, ART is particularly complex and time-consuming [11][12][13], generating increased workload for all treatment staff. As not all patients may benefit from ART, it is essential to identify early predictors of PG overdose to enable appropriate patient selection. Neck thickness, weight loss, PG volume, initial tumor volume, and decrease in tumor volume were found to correlate with PG overdose [10,14,15]. In a recent review [6], these anatomical parameters were identified as selection criteria for ART patient selection, though no clear conclusion was reached due to the heterogeneity of the studies. However, these correlations were mostly primarily identified using parameters calculated at the end of treatment, therefore significantly limiting the possibility of treatment modifications. These parameters were also correlated with each other. Decrease in PG volume, for instance, has been found to correlate with age, body mass index, planned dose to the parotid glands, initial PG volume, and the volume of PG overlapping with lymph node metastases. Decrease in PG volume may be a useful parameter for identifying patients at higher risk of xerostomia [27]. However, neither the Brouwer et al. [6] study nor our own found any clear association between the decrease in PG volume and PG overdose. Other parameters may be indirectly linked to PG overdose. Human papillomavirus (HPV)-positive cancer has demonstrated a higher sensitivity to radiation [28,29]. As the decrease in the CTV during the first week of treatment was correlated with PG overdose, HPV-status may exert an impact on the risk of PG overdose. Yet our series was not large enough to analyze the impact of this relationship.
Anatomical variations occurring during the first two weeks of treatment may be particularly relevant for PG overdose prediction, and may justify early replanning,  resulting in significant PG dose sparing [17,30]. Indeed, in the literature, CTV shrinkage appears to be particularly significant during the first week of radiotherapy [31]. In our study, the only anatomical parameters strongly correlated with PG overdose were CTV 70 at planning and its decrease in the first week. However, this parameter alone was not sufficient to predict final PG overdose. Indeed, the most relevant parameter in our nomogram was the PG dose difference between the planning CT (CT0) and the first week of treatment CT (CT1). Early PG overdose has also been demonstrated to correlate with the estimated cumulated PG dose by Hunter et al. [8].
In our study, we found that considering the parameters only for the planning CT was not sufficient for PG overdose prediction. The acquisition of a new CT, performed during the first week of treatment with dose calculation, was required to predict the mean cumulated PG dose at the end of treatment. In terms of practical use of ART, replanning decisions can be based solely on anatomical parameters, ideally defined on cone-beam CT (CBCT) performed at the time of the fraction, which is also useful for bone registration to correct for patient set-up. This approach should be explored further, assuming that anatomical parameters alone could be sufficient for PG overdose prediction, and also be visible on the CBCT.
We performed a leave-one-out cross validation to estimate the model's stability and accuracy, which revealed very low variation in each model coefficient. So, the model is not strongly influenced by individual patients, showing that, even if the number of patients is low, the The nomogram's sensitivity was only 80 %, which is insufficient for clinical decision making. The nomogram failed to predict PG overdose for two patients, whom exhibited moderate mean PG dose increases (<1.5Gy). In order to improve the sensitivity and identify other anatomical parameters correlated with PG overdose, a larger patient cohort is required. In addition, an external cohort is needed to validate the nomogram before it can be implemented in routine clinical practice.  [23,24] is represented on the right y-axis. Xerostomia was defined as a salivary flow ratio <25 % of the pretreatment one We attempted to translate the dosimetric benefit from ART to a clinical benefit, based on a xerostomia NTCP model. A strong relation existed between the mean doses to the parotid glands and salivary flow, and the efficacy of these NTCP models proved relatively high (area under the curve: 0.68-0.75) [24,[32][33][34]. Non-dosimetric patient factors, such as age, tumor stage, baseline xerostomia, and chemotherapy [32,[35][36][37] may also increase the risk of xerostomia, though this is controversial [32][33][34]. Using IMRT, severe xerostomia was detected in approximately 40 % of the patients, compared to 80 % using 3D conformal radiotherapy [38,39]. The dose constraints recently suggested by the QUANTEC group [34] may enable reducing severe xerostomia to under 20 % [33,37]. However, these dose constraints were only met in a minority of patients. In our study, the estimated xerostomia risk for the cumulated dose without replanning was 37 %, close to the observed xerostomia values in IMRT studies [38,39]. For the patients predicted to receive overdose by the nomogram (70 % of our population), an ART strategy could enable the xerostomia risk to be reduced to under 30 %.
The key issues affecting ART use include the choice of image registration method to monitor the cumulated dose and thus trigger replanning, within a doseguided radiotherapy perspective. The PGs are often close to or within a high-dose gradient, and even minor geometrical registration errors can lead to high cumulated-dose errors. A recent study evaluating accuracy for dose accumulation of 10 deformable-image registration methods in HNC [40]. For the most accurate method (FFD with mutual information), a mean cumulated dose point-to-point error of 2.5Gy for the PG was shown. Taking into account these uncertainties, it was possible to correctly identify the overall under-and overdosed PGs. Moreover, potential loss of PG cells during the course of radiotherapy [41] should be carefully considered due to the uncertainty of point-to-point matching between different CTs when using a deformable registration method. In addition, due to the possible heterogeneity of radiosensitivity within the PGs [42][43][44], local hotspots may also have different impacts on xerostomia risk. In our study, the predicted value of PG overdose was obtained from the estimated cumulated dose using the Dose Index Registry (DIR), and is therefore subject to uncertainty. For all these reasons, and in the interest of clinical justification (correlation between the mean PG dose and xerostomia), we only investigated the mean PG dose. Clearly, more thorough analysis into correlations between the PG cumulated dose and xerostomia, using a larger patient cohort or phantombased studies evaluating the accuracy of the DIR, need to be carried out in order to validate the dose accumulation method for daily clinical practice.

Conclusion
Based on the planning and first week CTs, our nomogram enabled the identification of all patients with large PG overdoses (≥2.5Gy). Replanning of these targeted patients lead to an eventual 4-Gy decrease in the mean PG overdose, while still respecting the dose-volume constraints in the other OARs and PTV. Other external cohorts are now required to validate this nomogram, as are clinical studies in order to validate the benefit of ART in the aims of decreasing the risk of xerostomia in locally-advanced HNC IMRT.