Introduction
Patients with head-and-neck cancer (HNC) are commonly treated with radiation therapy (RT). Of the spectrum of side effects owing to injury to organs at risk (OARs) from RT, xerostomia that arises from the parotid gland (PG) and submandibular gland (SMG) radiation have received significant study. The mean PG and SMG radiation dose is well known to be associated with the risk of developing xerostomia.
1- Eisbruch A.
- Kim H.M.
- Terrell J.E.
- et al.
Xerostomia and its predictors following parotid-sparing irradiation of head-and-neck cancer.
, 2- Dirix P.
- Nuyts S.
- Van Den Bogaert W.
Radiation-induced xerostomia in patients with head and neck cancer: A literature review.
, 3- Lee T.F.
- Liou M.H.
- Huang Y.J.
- et al.
LASSO NTCP predictors for the incidence of xerostomia in patients with head and neck squamous cell carcinoma and nasopharyngeal carcinoma.
, 4- Beetz I.
- Schilstra C.
- van der Schaaf A.
- et al.
NTCP models for patient-rated xerostomia and sticky saliva after treatment with intensity modulated radiotherapy for head and neck cancer: The role of dosimetric and clinical factors.
Modern intensity modulated RT techniques provide the opportunity to modify radiation dose distributions in PG and SGM and could potentially avoid radiation-induced xerostomia.
Despite these technological advancements, RT-induced xerostomia continues to be a significant clinical challenge that commonly affects patient-reported quality of life. To study RT-induced xerostomia, most existing literature used aggregated or summarized dosimetric predictors within certain organs, such as mean dose and dose-volume histogram (DVH) metrics in PG.
1- Eisbruch A.
- Kim H.M.
- Terrell J.E.
- et al.
Xerostomia and its predictors following parotid-sparing irradiation of head-and-neck cancer.
, 2- Dirix P.
- Nuyts S.
- Van Den Bogaert W.
Radiation-induced xerostomia in patients with head and neck cancer: A literature review.
, 3- Lee T.F.
- Liou M.H.
- Huang Y.J.
- et al.
LASSO NTCP predictors for the incidence of xerostomia in patients with head and neck squamous cell carcinoma and nasopharyngeal carcinoma.
, 5- Deasy J.O.
- Moiseenko V.
- Marks L.
- et al.
Radiotherapy dose-volume effects on salivary gland function.
The disadvantage of this modeling approach is that the spatial information for the radiation dose within an organ is lost. Different spatial distributions of a dose within an organ can yield the same mean dose and DVH metrics. A previous study by our group using DVH metrics showed that the level of low dose delivered to the combined PGs has a strong influence on xerostomia.
However, this previous study was not able to identify the spatial location of the influential regions using DVH metrics.
On the other hand, spatial information of the dose is key to understand the local dose effect on xerostomia. Preclinical investigations suggest that RT-induced xerostomia may be related to not only PG dosimetry, but also spatial location of secondary radiation damage within the PG.
7- Konings A.W.T.
- Faber H.
- Cotteleer F.
- et al.
Secondary radiation damage as the main cause for unexpected volume effects: A histopathologic study of the parotid gland.
, 8- Van Luijk P.
- Pringle S.
- Deasy J.O.
- et al.
Sparing the region of the salivary gland containing stem cells preserves saliva production after radiotherapy for head and neck cancer.
In rat models, irradiation of the caudal portion of the PG caused not only xerostomia, but was also associated with salivary function recovery in contrast with irradiation of the cranial portion. The investigators subsequently demonstrated that this recovery may be related to the presence of stem or progenitor cells that are responsible for the recovery of radiation-induced xerostomia.
9- Nanduri L.S.Y.
- Maimets M.
- Pringle S.A.
- et al.
Regeneration of irradiated salivary glands with stem cell marker expressing cells.
Data mining investigations by Robertson et al within their informatic infrastructure have demonstrated that the level of low-dose irradiation to PG is associated with xerostomia at 3 to 6 months.
10- Robertson S.P.
- Quon H.
- Kiess A.P.
- et al.
A data-mining framework for large scale analysis of dose-outcome relationships in a database of irradiated head and neck cancer patients.
, 11- Nakatsugawa M.
- Cheng Z.
- Goatman K.A.
- et al.
Radiomic analysis of salivary glands and its role for predicting xerostomia in irradiated head and neck cancer patients.
A pilot study by Quon et al of 30 patients has shown that the cranial half of the PG and its dosimetry may be more important in causing severe xerostomia in patients with HNC, and the current study sought to build on these observations.
12- Quon H.
- Park S.
- Plishker W.
- et al.
Preliminary clinical evidence of parotid subvolume radiosensitivity and the risk of radiation-induced xerostomia in head and neck cancer (HNC) patients.
To more robustly evaluate the influence of specific subvolumes of the PG and SMG on injury and symptoms of severe xerostomia after RT, we applied supervised machine learning methods and a radiomorphology approach using voxel dose to predict parotid-injury-causing acute xerostomia. Radiomorphology parametrically represents the spatial dose distribution within normalized anatomic structures using voxel- or shaped-based dosimetric predictors, which are consistent across patient cohorts.
13Lakshminarayanan P, Jiang W, Robertson S. Radio-morphology: parametric shape-based features in radiotherapy [published online December 3, 2018]. Med Phys. https://doi.org/10.1002/mp.13323.
The supervised machine learning method that was ultimately used is able to learn the influence of spatial dose patterns across organ subvolumes on xerostomia by assigning higher weights to the voxel dose of the predictive regions and lower weights to less predictive regions. We define the spatial pattern of the learned weights distributed across the voxels as the voxel importance pattern.
Discussion
The results demonstrate the successful application of the use of spatially explicit dosimetric predictors in predicting xerostomia in patients with HNC, leading to the identification of potential important parotid dose subvolumes associated with the predicted risk of severe xerostomia at 3 months after RT. To the best of our knowledge, this analysis used the largest cohort of patients with xerostomia HNC and prospective point-of-care data in literature, which enables robust modeling results and conclusions. The analysis confirmed evidence that parotid subvolumes are at a greater risk of contributing to radiation-induced xerostomia, as initially proposed in humans by van Luijk et al.
8- Van Luijk P.
- Pringle S.
- Deasy J.O.
- et al.
Sparing the region of the salivary gland containing stem cells preserves saliva production after radiotherapy for head and neck cancer.
Modeling the radiation dose spatially explicitly provides insights into the spatial dependence of dose in contrast with the use of DVH metrics alone. For instance, we are aware of 1 existing study that used a similar voxel-based dose analysis for acute dysphagia in patients with HNC.
22- Monti S.
- Palma G.
- D'Avino V.
- et al.
Voxel-based analysis unveils regional dose differences associated with radiation-induced morbidity in head and neck cancer patients.
The authors applied a statistical test to compare the voxel-based dose distribution between patients with grade ≥3 versus <3 dysphagia. Instead of comparing the dose difference between patients who developed xerostomia and those who did not, we combined the voxel-based dose modeling with supervised machine learning algorithms to build a predictive model for xerostomia. We also learned how spatial dose influences xerostomia, highlighting the flexibility of this methodology. Our methodology can also be potentially applied to other RT-related head-and-neck toxicities, such as dysphagia and trismus.
The distribution of average doses shows that the most important region found (ie, superior-anterior portion of the contralateral PG) is within the low-dose region in the patient cohort. This is consistent with the results from a study by Robertson et al using DVH metrics to demonstrate the impact of a low-dose bath to both PG on the risk of grade ≥2 xerostomia.
10- Robertson S.P.
- Quon H.
- Kiess A.P.
- et al.
A data-mining framework for large scale analysis of dose-outcome relationships in a database of irradiated head and neck cancer patients.
We believe that if this region was treated with a high radiation dose, the other regions were likely treated with an even higher dose, leading to a much higher risk of developing xerostomia.
Furthermore, we would expect the superior portion of the ipsilateral PG also to be more predictive because the region is also low dose. However, the voxel importance pattern result did not reveal this, but rather, the medial portion of the ipsilateral PG was more predictive. We hypothesize that in the ipsilateral PG, the medial portion is more important because of its proximity to the ductal region, and radiation damage to that region will render doses in the superior portion much less important in predicting xerostomia. This is in accordance with the preclinical xerostomia model that has demonstrated that sparing the stem or progenitor cells contained in the ductal region of PG preserves salivary function after RT.
8- Van Luijk P.
- Pringle S.
- Deasy J.O.
- et al.
Sparing the region of the salivary gland containing stem cells preserves saliva production after radiotherapy for head and neck cancer.
We further investigated whether our current method improves the prediction performance compared with using only conventional dosimetric features. Specifically, we computed the mean dose and DVH features (D5, D25, D50, and D90) over all parotid and submandibular glands. The AUC performance using these conventional dosimetric features and the ridge logistic regression model evaluated with nested 10-fold cross-validation on the same 477 patients is 0.68 ± 0.09, which is not significantly different from using voxel-dose features. Although the current method using voxel-dose features did not significantly improve prediction performance, more insight is provided about the spatial dependency between radiation dose and acute xerostomia, which cannot be achieved using only conventional dosimetric features.
The findings provide insight into the importance of bilateral parotid injury,and underscore the importance of carefully determining the clinical indications for bilateral cervical nodal irradiation along with a careful delineation of how superior and lateral the cervical nodal planning target volumes encroach upon the medial and superior aspects of the PG. Moreover, the analysis also demonstrated (to a lesser degree) the importance of irradiation in the subvolumes in the contralateral SMG contributing to the severity of xerostomia at 3 months post-RT. This highlights the clinical implications and importance of reducing the volume of cervical neck irradiation as a clinical strategy to de-intensify the current chemoradiation treatment paradigm for head-and-neck squamous cell carcinomas, especially for HPV-associated oropharyngeal carcinomas.
Several additional limitations of our analysis need to be recognized. First, the study population consists of patients treated at a single local hospital. The specific voxel importance pattern obtained may be specific for patients at this hospital. Second, for the xerostomia prediction outcomes, we used last observation carried forward to obtain data for patients without xerostomia assessment at 3 months post-RT. Our longitudinal xerostomia outcomes data show that there is minor xerostomia recovery from the end of treatment to 3 months post-RT. Therefore, our prediction outcome data could be slightly biased toward a more severe xerostomia grade.
Third, we believe that the importance pattern identified by our method depends not only on the actual relationship between dose and xerostomia outcome, but also on the range of dose variations in our patient cohort. Most patients have similar patterns of dose; therefore, evaluating dose-response outside of the range of patterns delivered to patients in the database is challenging. Even if we think that areas with low-dose variation are likely to show weak associations, our study still shows that the low-dose region on the contralateral parotid gland, which has a low-dose variation, is the most influential region.
Fourth, the learned influence pattern depends on the specific model we used (ie, ridge logistic regression). The identified important region may be too big if the region that contains highly correlated voxel dose is large. In addition, our results indicate a possible serial component to the behavior within the ipsilateral parotid (ie, high radiation dose delivered to medial portion of ipsilateral parotid [close to ductal region]), and possibly causing an occlusion to the duct, rendering the superior portion less important.
However, we did not explicitly model serial behavior within organs because we modeled voxel dose as independent features. This assumes that doses in different regions within the analyzed organs influence xerostomia in parallel. Explicitly modeling the dependency between doses in different voxels may enable us to model possible serial behavior of dose influence on xerostomia within organs. However, we believe a hypothesis regarding certain serial behavior or dependency between dose influence on xerostomia in different regions is required beforehand to investigate possible serial behavior. Without prior knowledge and a hypothesis about any serial behavior, we adopted the current modeling approach.
Fifth, the pattern of dose effect on xerostomia, as we learned in this study, represents only the association between radiation dose and xerostomia. Sixth, because we treated each voxel independently, we did not explicitly consider the spatial coherence of the dose patterns. However, ridge logistic regression handles the correlations between voxel doses by assigning similar voxel importance to highly correlated predictors, which is inherent in the way we treated patients. Future efforts that explicitly include the spatial coherence between dose voxels in derived features may uncover interdependencies that also relate to the xerostomia outcome. No conclusion on the causal effect between dose and xerostomia was established in this observational study. Finally, this study only included radiation doses in the PG and SMG, but there is a study that reported that the mean dose to oral cavity is associated with xerostomia as well.
1- Eisbruch A.
- Kim H.M.
- Terrell J.E.
- et al.
Xerostomia and its predictors following parotid-sparing irradiation of head-and-neck cancer.
For future studies, we can apply this approach to patients with HNC at other hospitals that may have different patient characteristics and radiation treatment plans to validate our approach and the voxel importance pattern we learned. Ultimately, we want to learn the causal effect between radiation dose and xerostomia, which is difficult using observational studies. Either an experimental study or advanced causal inference analysis should be conducted. Nabi and Shpitser recently proposed a causal inference technique (ie, causal sufficient dimension reduction) for high-dimensional treatments problems,
23Semi-parametric causal sufficient dimension reduction of high dimensional treatments.
which we are currently applying to our problem.
In this study, we focused on the methodology to study the spatial pattern of dose influence on xerostomia. To investigate the effect of possible interaction, reserve capacity, and compensation mechanism on xerostomia, we are currently studying the spatial dose influence on xerostomia recovery by comparing its influence pattern with an acute xerostomia analysis. We believe this next study will provide further implications of the biological mechanisms behind dose effect on xerostomia.
Different dose and fractionation schemes are known to have different biologic effectiveness for HNC treatment. For our patient cohort, all 427 patients were treated with 33 to 36 fractions, and the highest dose-planning target volume received was between 200 Gy and 220 Gy. Therefore, fractionation schemes do not vary much among our patient cohort. However, we believe a study of the effect of different fractionation schemes on xerostomia would be interesting for future studies.
In addition, whether the current method combined with conventional dosimetric features will yield better prediction accuracy is another interesting research question for future study. To combine the current method with conventional dosimetric features, we can carve the parotid and submandibular glands into different subregions (larger than voxels), and compute dosimetric features for each of the subregions. Those dosimetric features within the carved subregions can be used to predict xerostomia. Using this approach, we will still be able to capture the spatial dependency between radiation dose and xerostomia and enables a more flexible modeling approach because we can control the sizes of the carved subregions, which may improve xerostomia prediction accuracy. Finally, we will also include dose in oral cavity in our study to investigate how dose in subvolumes in oral cavity is associated with xerostomia.
Article info
Publication history
Published online: November 28, 2018
Accepted:
November 14,
2018
Received:
August 23,
2018
Footnotes
Sources of support: This work was funded by the Radiation Oncology Institute.
Conflicts of interest: All authors have no potential conflict of interest to disclose, except for the following 3 authors: Dr. McNutt reports grants from the Radiation Oncology Institute during the conduct of the study and grants from Canon Medical and Philips Health Care outside the submitted work. In addition, Dr. McNutt holds a pending patent for “Method and apparatus for determining treatment region and mitigating radiation toxicity” (20170259083) and patents for “Method, system, and computer-readable media for treatment plan risk analysis” and “System and method for medical data analysis and sharing” (20170083682 and 20160378919, respectively). Dr. Taylor holds US patent 8,688,618 B2, “Shape-based retrieval of prior patients for automation and quality control of radiation therapy treatment Plans” (filed June 22, 2009; issued April 1, 2014, with royalties paid to unknown). Mr. Bowers reports grants from Elekta during the conduct of the study.
Copyright
© 2018 The Authors. Published by Elsevier Inc. on behalf of American Society for Radiation Oncology.