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

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 [2‐5], the percentage of patients with estimated PG overdoses ranges widely from 5 to 70 % [1, 5‐10]. 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 [11‐13] 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 over-irradiation [10, 14‐17]. 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.

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):

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 (∆DosePG) (r² = 0.5 and 0.9, respectively, p <0.001), only the parameter with the highest r² was used (mean PG dose difference). The three parameters used for PG overdose prediction were CTV₇₀ on the planning CT dataset (CTV_{70_CT0}), the CTV₇₀ CT0-to-CT1 difference (∆CTV₇₀), and the difference between the mean PG doses of CT1 and CT0 (∆DosePG).

Based on these parameters, the resulting linear regression model for PG overdose prediction, optimized for all patients, was:

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 (r² = 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 leave-one-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 \( \mathrm{C}\mathrm{T}\mathrm{V}{70}_{\mathrm{CT}0} \) _, \( \Delta CTV70 \) _, and \( \Delta DosePG \), 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 %).

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 [11‐13], 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 CTV₇₀ 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 considered population is homogeneous enough for the model to reach stability. No outliers were extracted. These results are proof of the model’s good stability and accuracy.

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.

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, 32‐34]. Non-dosimetric patient factors, such as age, tumor stage, baseline xerostomia, and chemotherapy [32, 35‐37] may also increase the risk of xerostomia, though this is controversial [32‐34]. 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 dose-guided 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 [42‐44], 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 phantom-based studies evaluating the accuracy of the DIR, need to be carried out in order to validate the dose accumulation method for daily clinical practice.

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.

∆CTV₇₀, CTV₇₀ CT0-to-CT1 difference; ∆DosePG, difference between the mean PG doses of CT1 and CT0; ART, Adaptive Radiotherapy; CBCT, Cone-beam CT; CTV, Clinical Target Volume; FFD: free-form deformation; GTV, Gross Tumor Volume; LAHNC, locally-advanced head and neck carcinoma; NTCP, normal tissue complication probability; OAR, Organ at risk; PG, Parotid gland; PTV, Planning Target Volume.